CN102436576B - Multi-scale self-adaptive high-efficiency target image identification method based on multi-level structure - Google Patents
Multi-scale self-adaptive high-efficiency target image identification method based on multi-level structure Download PDFInfo
- Publication number
- CN102436576B CN102436576B CN201110321561.8A CN201110321561A CN102436576B CN 102436576 B CN102436576 B CN 102436576B CN 201110321561 A CN201110321561 A CN 201110321561A CN 102436576 B CN102436576 B CN 102436576B
- Authority
- CN
- China
- Prior art keywords
- thumbnail
- target image
- image
- layer
- far infrared
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Expired - Fee Related
Links
- 238000000034 method Methods 0.000 title claims abstract description 76
- 238000000354 decomposition reaction Methods 0.000 claims abstract description 8
- 239000010410 layer Substances 0.000 claims description 114
- 238000004088 simulation Methods 0.000 claims description 35
- 238000006243 chemical reaction Methods 0.000 claims description 16
- 238000012360 testing method Methods 0.000 claims description 15
- 238000012549 training Methods 0.000 claims description 8
- 239000011229 interlayer Substances 0.000 claims description 7
- 230000003595 spectral effect Effects 0.000 claims description 7
- 238000001228 spectrum Methods 0.000 claims description 7
- 230000009471 action Effects 0.000 claims description 5
- PGLTVOMIXTUURA-UHFFFAOYSA-N iodoacetamide Chemical compound NC(=O)CI PGLTVOMIXTUURA-UHFFFAOYSA-N 0.000 claims description 5
- 238000012856 packing Methods 0.000 claims description 5
- 230000001360 synchronised effect Effects 0.000 claims description 5
- 230000008859 change Effects 0.000 abstract description 2
- 210000000056 organ Anatomy 0.000 description 7
- 238000004458 analytical method Methods 0.000 description 4
- 230000000694 effects Effects 0.000 description 4
- 210000001328 optic nerve Anatomy 0.000 description 4
- 238000012545 processing Methods 0.000 description 4
- 230000008878 coupling Effects 0.000 description 3
- 238000010168 coupling process Methods 0.000 description 3
- 238000005859 coupling reaction Methods 0.000 description 3
- 238000005286 illumination Methods 0.000 description 3
- 210000000857 visual cortex Anatomy 0.000 description 3
- 241000282320 Panthera leo Species 0.000 description 2
- 230000007812 deficiency Effects 0.000 description 2
- 230000002950 deficient Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 210000004709 eyebrow Anatomy 0.000 description 2
- 210000000887 face Anatomy 0.000 description 2
- 238000001914 filtration Methods 0.000 description 2
- 230000000007 visual effect Effects 0.000 description 2
- 240000008042 Zea mays Species 0.000 description 1
- 235000005824 Zea mays ssp. parviglumis Nutrition 0.000 description 1
- 235000002017 Zea mays subsp mays Nutrition 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 230000000903 blocking effect Effects 0.000 description 1
- 235000005822 corn Nutrition 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000003708 edge detection Methods 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 239000011521 glass Substances 0.000 description 1
- 210000004209 hair Anatomy 0.000 description 1
- 238000010191 image analysis Methods 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 230000013011 mating Effects 0.000 description 1
- 210000000653 nervous system Anatomy 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000002360 preparation method Methods 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 238000012706 support-vector machine Methods 0.000 description 1
Landscapes
- Image Analysis (AREA)
Abstract
The invention relates to a multi-scale self-adaptive high-efficiency target image identification method based on a multi-level structure. The method comprises the steps of acquiring a target image; performing multi-layer zooming, decomposition and display on the acquired original image and identifying the target image. By the method, the problems of instability of an identification system, low corresponding rate of the relationship between the identification accuracy rate and the identification speed, low identification rate and the like caused by factors such as size change of the image and the like in image identification are completely solved. In addition, the multi-scale self-adaptive high-efficiency target image identification method based on the multi-level structure has high identification speed.
Description
Technical field
The present invention relates to a kind of object features image-recognizing method, especially relate to a kind of multi-scale self-adaptive high-efficiency target image identification method based on the multi-layer structure.
Background technology
How from can identifying specific objective as human eye image, for example: desk, automobile etc. are mankind's dream and pursuits of one always.Image object identification is the major issue of a research of artificial intelligence field, is solving automated production and detection, productive life the very corn of a subject methods such as intelligent image analysis and retrieval.
Present stage, target accurately location is the gordian technique that identifying information is processed, and is widely used in the systems such as recognition of face, man-machine interaction, Intelligent Human-Machine Interface.Under complex background, the accurate location of eyes is also one and has challenging problem.This is except being because the factors such as illumination, size, attitude, Plane Rotation, picture quality are brought complicated variation to the eyes outward appearance, and blocking etc. of reflective, the hair of the switching of eyes, glasses and picture frame also brings a lot of difficulties to the accurate location of eyes; Particularly in the situation that eyes closed, eyebrow and thick picture frame all can bring larger interference to the eyes location.
The eyes that propose are at present accurately located the method that main stream approach is based on heuristic rule.These class methods are mainly to formulate locating rule according to the priori of eyes.These prioris comprise organ distributed knowledge, shape knowledge, color knowledge, physical characteristics etc.It is relatively poor that these class methods generally adapt to the extraneous ability that changes, and one or more that often can only treatment of organs change, and the requirement of stability and precision and practical application also has gap.Causing the reason of this phenomenon is mainly the local appearance that they have only considered organ, and does not consider the restriction relation between organ and adjacent domain or organ.When there be the object similar to the target organ outward appearance in face, will affect to positioning belt like this.Outward appearance during such as eyes closed and eyebrow, thick picture frame are very similar.So the global characteristics that considers the organ local feature and can express this restriction relation could obtain the more locating effect of robust.
Along with the progress of technology, in target identification, how rapidly and efficiently the extraction target signature and to set up corresponding identification system be an important and crucial problem in target identification.In Target Recognition Algorithms commonly used, template matches (template matching) is a kind of important method at present.Stencil matching is by adopting masterplate that the method that image carries out global registration is identified target.What have is a little that algorithm is simple, easily realizes.Shortcoming is that the robustness (Robustness) of algorithm is poor especially, size when image, when directions etc. have conversion, the non-constant of matching effect, although someone proposes to adopt the method for the masterplate of a plurality of different size and Orientations to solve problems, but this solution has increased the time complexity of algorithm greatly, and making becomes in the middle of practical application hardly may.David Marr, Cannon etc. have proposed the method that different edge extracting (edge detection) is another target identification, it has is a little the geometric properties of response diagram picture to a certain extent, has certain adaptability to light and shade conversion etc. is arranged.Recently David lowe is at " Objectrecognitionfrom local scale-invariant features ", and " SIFT " algorithm that proposes in documents such as " Distinctiveimage features from scale-invariant keypoints " is more and more paid attention to and uses in target identification.Then the SIFT algorithm extracts the method for scale-invaraint feature and identifies by adopting Gaussian filter bank to carry out filtering to image.This method is to specific objective, for example: certain specific books, the identification aspect of building etc. has obtained effect preferably.Yet the SIFT algorithm is in the object identification to a class, for example: picking out all desks and the automobile texts is unsatisfactory from image, is the combination of serial of methods on SIFT algorithm essence in addition, and the time complexity of algorithm is very high.And in document " Object recognition from local scale-invariant features ", author David lowe mentions, and the SIFT algorithm just meets the optic nerve principle of human body to a certain extent.Because the image object that will identify such as may exist between size, direction, illumination, deformation, class at the variation, image recognition is a very complicated problem.Generally speaking, due to the complicacy of target identification problem, existing method can well not felt above problem aspect the identification accuracy of method and speed.Exist especially:
(1) size of image, direction, illumination, individual etc. vicissitudinous the time, the identification target of robustness how.
(2) how can reasonably can identify fast target in time range.
Deficiency or defective for present recognition methods, the present invention proposes a kind of brand-new multiple dimensioned efficient target identification method with hierarchical structure: mainly comprise following which floor: at first, analysis tool (wavelet analysis or pyramid multiscale analysis method) with Multi-scale, image is carried out filtering to be processed, obtain the decomposition texture of a level, self-adaptation is chosen matching layer and is mated, obtain the secondary characteristics of image, based on the validity feature of this extraction, can carry out high efficiency target identification.
Summary of the invention
The objective of the invention is for image identification system deficiency of the prior art and defective, a kind of effectively rapidly and efficiently image-recognizing method, specifically a kind of multi-scale self-adaptive high-efficiency target image identification method based on the multi-layer structure are provided.Multiple dimensioned (multi-scale) image object recognition methods that has concrete hierarchical structure (Hierarchical) by employing, the factors such as size conversion due to image of more comprehensively having solved in image recognition cause that recognition system is stable not, concern the problems such as corresponding rate is lower, discrimination is not high between recognition accuracy and recognition speed, and the multi-scale self-adaptive high-efficiency target image identification method based on the multi-layer structure of the present invention also has very high recognition speed simultaneously.
For achieving the above object, the present invention is achieved by the following technical solutions:
A kind of multi-scale self-adaptive high-efficiency target image identification method based on the multi-layer structure of the present invention is characterized in that, should comprise the following steps successively based on the multiple dimensioned high-efficiency target image identification method of multi-layer structure:
Step 1, collection target image; The acquisition method of described target image is as follows:
At first, be separately installed with initiatively far infrared light source on the far infrared digital simulation field camera of the subsidiary image pick-up card of a plurality of outputs; Wherein, select the far red light spectral coverage of whole spectrum medium wavelength between 9um and 11um as active far infrared light source, initiatively the far infrared light source is comprised of the N of wavelength between a 9.5um and 10.5um LED, and wherein N is natural number, N ∈ [26,36]; When the far infrared digital simulation field camera of the described active far infrared light source of combination and the subsidiary image pick-up card of described output, with the coaxial arrangement of far infrared digital simulation field camera of LED and the subsidiary image pick-up card of described output; The be del shape of described light emitting diode in the far infrared digital simulation field camera plane of the subsidiary image pick-up card of described output evenly put, with the light source stepless action in target image;
Collection for target image, use the target site in the visible light digital simulation field camera alignment image of the subsidiary image pick-up card of a described output to gather, and by image capture software, the far infrared digital simulation field camera of the visible light digital simulation field camera that is connected to the subsidiary image pick-up card of above-mentioned output on the same image pick-up card and the subsidiary image pick-up card of above-mentioned output is carried out synchronous acquisition and control, image acquisition has simultaneity;
The original image multilayer convergent-divergent of step 2, collection decomposes and shows; The original image convergent-divergent decomposition display packing of described collection is as follows: generate the target image thumbnail data according to the image acquisition data in step 1, the thumbnail data data type is the two-dimensional array that and display device show lattice match; Then generate corresponding thumbnail according to described target image thumbnail data, this is generated corresponding thumbnail, carry out turriform anisotropic filter group conversion more than 5 layers; Again the thumbnail of described bank of filters conversion rearranged the layering classification, thumbnail layer corresponding to del top in above-mentioned steps one is classified as S
1Layer, thumbnail layer corresponding to del bottom in step 1 is classified as S
nLayer, described S
1Layer and S
nContain some layers between layer, its S
nIn n be natural number, n ∈ [34,39]; Wherein, adjacent two thumbnail interlayer thumbnail image scale size ratios are 1～2: 1, and the thumbnail image of all thumbnail layers forms pyramid;
Step 3, target image identification; Described target image identification method is as follows: adopt linear SVM to classify to all the thumbnail layers in the pyramid in above-mentioned steps two, when classifying, choose 2/5ths to 3/5ths thumbnail layer as training set, then remaining thumbnail layer carries out the multiple dimensioned high-efficiency target image identification of multi-layer structure as test set as the standard of test accuracy by the equilibrium point on the demonstration curve that generates in described linear SVM.
As preferred technical scheme:
Adjacent two thumbnail interlayer thumbnail image scale size ratios in above-mentioned steps two are 1.414: 1.
S in above-mentioned steps two
1Layer and S
nContain 36 layers between layer.
The thumbnail corresponding to this generation in above-mentioned steps two carries out the conversion of 16 layers of turriform anisotropic filter group.
The thumbnail layer of choosing half in above-mentioned steps three is as training set.
The far red light spectral coverage that the whole spectrum medium wavelength of selection in above-mentioned steps one is 10um is as active far infrared light source.
N the LED that active far infrared light source in above-mentioned steps one is 10.5um by wavelength forms.
The present invention is by researching and analysing the relative merits of image identification system in the past, various challenges and problem that analysis image identification faces, consider recognition accuracy in image recognition and the relation between recognition speed, multiple dimensioned (multi-scale) image object recognition methods that has concrete hierarchical structure (Hierarchical) by employing, more comprehensively having solved in the factors such as size conversion due to image in image recognition causes recognition system stable not, the problems such as discrimination is not high, the method has very high recognition speed simultaneously.
Simultaneously, the system that whole recognition methods of the present invention consists of adopts the Hierarchical hierarchy that meets human visual nervous system (Visual cortex), can also be by S1, and C1, S2, C2 four is large, and layer consists of.The method of every large layer is described as follows:
S1: in this layer, the method that can adopt Multi-scale to analyze is decomposed image, then the image that decomposes is carried out obtaining the Multi-resolution processing scheme of pyramid from group.
After adopting the method for Multi-scale to decompose to image, can obtain the structure of a pyramid.Usually, in the structure that this kind method obtains, the image tool has plenty of the decay of 2: 1 sizes, and namely between adjacent two layers, the image size is 2: 1.For make image layer and layer directly a size be arbitrary proportion, for example 1.414: 1, and traditional method adopts the method for interpolation to obtain intermediate image, have to a certain degree distortion but the method by interpolation obtains image, computing velocity is slow simultaneously.
At this, we propose a kind ofly in advance image to be carried out convergent-divergent (rescale), and then carry out multi-scale and decompose (wavelet decomposition etc.), then the image after decomposing are rearranged (Reorder), then obtain the S1 layer of image.So just can obtain any band, the size between band is the method for the structure of arbitrary proportion.
Complex cell (Complex cell) in C1:C1 layer main phase pair and optic nerve (visual cortex).By to merger many between adjacent layer, obtain the c1 layer of image.
The S2:S2 layer is equivalent to V2 in optic nerve (Visual Cortex) and the simple cell in the v4 zone.And at this layer, can carry out adaptively choosing different band and mating.In the middle of this layer, a selected k feature and former masterplate feature are carried out the coupling of following formula,
C2: among this layer, adopt the operation of global maximum, to each masterplate feature, obtain an eigenwert correspondingly.The feature that obtains is the biological characteristic that meets human body optic nerve system.
Classification: then adopt sorter that the eigenwert of obtaining is classified, realize target identification.
recognition methods of the present invention has above by adopting the Multi-scale model of level, having avoided traditional template matches target is method to be needed the method for the template of a plurality of different scales, and by the adaptive method of choosing coupling level (Band), realized the high-level efficiency of coupling, and whole target recognition structure matches with people's visual system to a certain extent, the robustness of the recognizer that exists and the problem of the contradiction between the algorithm time complexity have more comprehensively been solved in field of target recognition, experimental result shows that the method has good recognition accuracy and has simultaneously higher recognition speed.
Embodiment
Below in conjunction with embodiment, further set forth the present invention.
Embodiment 1:
A kind of multi-scale self-adaptive high-efficiency target image identification method based on the multi-layer structure is characterized in that, should comprise the following steps successively based on the multiple dimensioned high-efficiency target image identification method of multi-layer structure:
Step 1, collection target image; The acquisition method of described target image is as follows:
At first, be separately installed with initiatively far infrared light source on the far infrared digital simulation field camera of the subsidiary image pick-up card of a plurality of outputs; Wherein, select the far red light spectral coverage of whole spectrum medium wavelength between 9um and 11um as active far infrared light source, initiatively the far infrared light source is comprised of the N of wavelength between a 9.5um and 10.5um LED, and wherein N is natural number, N ∈ [26,36]; When the far infrared digital simulation field camera of the described active far infrared light source of combination and the subsidiary image pick-up card of described output, with the coaxial arrangement of far infrared digital simulation field camera of LED and the subsidiary image pick-up card of described output; The be del shape of described light emitting diode in the far infrared digital simulation field camera plane of the subsidiary image pick-up card of described output evenly put, with the light source stepless action in target image;
Collection for target image, use the target site in the visible light digital simulation field camera alignment image of the subsidiary image pick-up card of a described output to gather, and by image capture software, the far infrared digital simulation field camera of the visible light digital simulation field camera that is connected to the subsidiary image pick-up card of above-mentioned output on the same image pick-up card and the subsidiary image pick-up card of above-mentioned output is carried out synchronous acquisition and control, image acquisition has simultaneity;
The original image multilayer convergent-divergent of step 2, collection decomposes and shows; The original image convergent-divergent decomposition display packing of described collection is as follows: generate the target image thumbnail data according to the image acquisition data in step 1, the thumbnail data data type is the two-dimensional array that and display device show lattice match; Then generate corresponding thumbnail according to described target image thumbnail data, this is generated corresponding thumbnail, carry out turriform anisotropic filter group conversion more than 16 layers; Again the thumbnail of described bank of filters conversion rearranged the layering classification, thumbnail layer corresponding to del top in above-mentioned steps one is classified as S
1Layer, thumbnail layer corresponding to del bottom in step 1 is classified as S
nLayer, described S
1Layer and S
nContain some layers between layer, its S
nIn n be natural number, n ∈ [34,36]; Wherein, adjacent two thumbnail interlayer thumbnail image scale size ratios are 1: 1, and the thumbnail image of all thumbnail layers forms pyramid;
Step 3, target image identification; Described target image identification method is as follows: adopt linear SVM to classify to all the thumbnail layers in the pyramid in above-mentioned steps two, when classifying, choose 2/5ths thumbnail layer as training set, then remaining thumbnail layer carries out the multiple dimensioned high-efficiency target image identification of multi-layer structure as test set as the standard of test accuracy by the equilibrium point on the demonstration curve that generates in described linear SVM.
Embodiment 2:
A kind of multi-scale self-adaptive high-efficiency target image identification method based on the multi-layer structure is characterized in that, should comprise the following steps successively based on the multiple dimensioned high-efficiency target image identification method of multi-layer structure:
Step 1, collection target image; The acquisition method of described target image is as follows:
At first, be separately installed with initiatively far infrared light source on the far infrared digital simulation field camera of the subsidiary image pick-up card of a plurality of outputs; Wherein, select the far red light spectral coverage of whole spectrum medium wavelength between 9um and 11um as active far infrared light source, initiatively the far infrared light source is comprised of the N of wavelength between a 9.5um and 10.5um LED, and wherein N is natural number, N ∈ [26,36]; When the far infrared digital simulation field camera of the described active far infrared light source of combination and the subsidiary image pick-up card of described output, with the coaxial arrangement of far infrared digital simulation field camera of LED and the subsidiary image pick-up card of described output; The be del shape of described light emitting diode in the far infrared digital simulation field camera plane of the subsidiary image pick-up card of described output evenly put, with the light source stepless action in target image;
Collection for target image, use the target site in the visible light digital simulation field camera alignment image of the subsidiary image pick-up card of a described output to gather, and by image capture software, the far infrared digital simulation field camera of the visible light digital simulation field camera that is connected to the subsidiary image pick-up card of above-mentioned output on the same image pick-up card and the subsidiary image pick-up card of above-mentioned output is carried out synchronous acquisition and control, image acquisition has simultaneity;
The original image multilayer convergent-divergent of step 2, collection decomposes and shows; The original image convergent-divergent decomposition display packing of described collection is as follows: generate the target image thumbnail data according to the image acquisition data in step 1, the thumbnail data data type is the two-dimensional array that and display device show lattice match; Then generate corresponding thumbnail according to described target image thumbnail data, this is generated corresponding thumbnail, carry out turriform anisotropic filter group conversion more than 16 layers; Again the thumbnail of described bank of filters conversion rearranged the layering classification, thumbnail layer corresponding to del top in above-mentioned steps one is classified as S
1Layer, thumbnail layer corresponding to del bottom in step 1 is classified as S
nLayer, described S
1Layer and S
nContain some layers between layer, its S
nIn n be natural number, n ∈ [36,39]; Wherein, adjacent two thumbnail interlayer thumbnail image scale size ratios are 1.414: 1, and the thumbnail image of all thumbnail layers forms pyramid;
Step 3, target image identification; Described target image identification method is as follows: adopt linear SVM to classify to all the thumbnail layers in the pyramid in above-mentioned steps two, when classifying, choose the thumbnail layer of half as training set, then remaining thumbnail layer carries out the multiple dimensioned high-efficiency target image identification of multi-layer structure as test set as the standard of test accuracy by the equilibrium point on the demonstration curve that generates in described linear SVM.
Embodiment 3:
A kind of multi-scale self-adaptive high-efficiency target image identification method based on the multi-layer structure is characterized in that, should comprise the following steps successively based on the multiple dimensioned high-efficiency target image identification method of multi-layer structure:
Step 1, collection target image; The acquisition method of described target image is as follows:
At first, be separately installed with initiatively far infrared light source on the far infrared digital simulation field camera of the subsidiary image pick-up card of a plurality of outputs; Wherein, select whole spectrum medium wavelength be far red light spectral coverage between 10um as active far infrared light source, initiatively the far infrared light source is that N LED between 10.5um forms by wavelength, wherein N is natural number, N ∈ [26,36]; When the far infrared digital simulation field camera of the described active far infrared light source of combination and the subsidiary image pick-up card of described output, with the coaxial arrangement of far infrared digital simulation field camera of LED and the subsidiary image pick-up card of described output; The be del shape of described light emitting diode in the far infrared digital simulation field camera plane of the subsidiary image pick-up card of described output evenly put, with the light source stepless action in target image;
Collection for target image, use the target site in the visible light digital simulation field camera alignment image of the subsidiary image pick-up card of a described output to gather, and by image capture software, the far infrared digital simulation field camera of the visible light digital simulation field camera that is connected to the subsidiary image pick-up card of above-mentioned output on the same image pick-up card and the subsidiary image pick-up card of above-mentioned output is carried out synchronous acquisition and control, image acquisition has simultaneity;
The original image multilayer convergent-divergent of step 2, collection decomposes and shows; The original image convergent-divergent decomposition display packing of described collection is as follows: generate the target image thumbnail data according to the image acquisition data in step 1, the thumbnail data data type is the two-dimensional array that and display device show lattice match; Then generate corresponding thumbnail according to described target image thumbnail data, this is generated corresponding thumbnail, carry out turriform anisotropic filter group conversion more than 8 layers; Again the thumbnail of described bank of filters conversion rearranged the layering classification, thumbnail layer corresponding to del top in above-mentioned steps one is classified as S
1Layer, thumbnail layer corresponding to del bottom in step 1 is classified as S
nLayer, described S
1Layer and S
nContain some layers between layer, its S
nIn n be natural number, n ∈ [36,39]; Wherein, adjacent two thumbnail interlayer thumbnail image scale size ratios are 1.414: 1, and the thumbnail image of all thumbnail layers forms pyramid;
Step 3, target image identification; Described target image identification method is as follows: adopt linear SVM to classify to all the thumbnail layers in the pyramid in above-mentioned steps two, when classifying, choose 3/5ths thumbnail layer as training set, then remaining thumbnail layer carries out the multiple dimensioned high-efficiency target image identification of multi-layer structure as test set as the standard of test accuracy by the equilibrium point on the demonstration curve that generates in described linear SVM.
Embodiment 4:
Can adopt the equipment such as camera and video camera to obtain required image, then the image that obtains be identified.
For method is carried out objective appraisal, adopted standard database that method is tested, database comprises 186 leaves, 1155 automobiles, 450 people's faces, 1074 Zhang Fei's machines, 826 motorcycles and 900 background pictures consist of.At the S1 layer, we adopt 4 band, and each band has 2 scale, and the scale between two adjacent scale differs 1.414, then according to recombinating as the mode of figure Fig.2.When classifying, choose the feature of 1000 C2 layers as test feature, adopt linear SVM (Support Vector Machine, SVM) classify. when classifying, we choose the image of half as training set, in addition half as test set, then seek equilibrium point on the RoC curve as the standard of test accuracy.
When testing, the hardware environment that adopts is: Intel Core2, and 6600 2.4GHz CPU, 3.24GB memory, 32bit Matlab 2007a,
We can find out from above-mentioned test, and after adopting new method, the preparation rate of identification all has lifting in various degree, is not particularly the targets such as good leaf and automobile to former recognition effect, and the accuracy rate of identification has the raising of significance degree.And the identification degree of readiness of all targets has all been reached the high discrimination more than 96%, satisfy most of identification personage's requirement.
Another advantage of algorithm is that recognition speed is very fast, the above-mentioned 5 classification target data that comprise are approximately tested in 5000 pictures storehouses, in the situation that adopt the matlab platform and do not carry out hardware optimization, the processing time of average every pictures was less than for 2 seconds, whole system is expected to reach the requirement of real time processing system, at a high speed shape library and video is implemented to process.Following table is the average handling time of every kind of target:
| The target classification | Leaf | Automobile | People's face | Aircraft | Motorcycle |
| Time | 1.8 second | 1.5 second | 1.9 second | 1.6 second | 1.8 second |
In average every pictures processing time of upper table, be total to approximately 5000 pictures, the very potential application requirements that reaches.
The present invention is not limited to top explanation and embodiment.On the contrary, being intended to the present invention extensively is suitable in the determined boundary of described claim below.
Claims (9)
1. the multi-scale self-adaptive high-efficiency target image identification method based on the multi-layer structure, is characterized in that, should comprise the following steps successively based on the multiple dimensioned high-efficiency target image identification method of multi-layer structure:
Step 1, collection target image; The acquisition method of described target image is as follows:
At first, be separately installed with initiatively far infrared light source on the far infrared digital simulation field camera of the subsidiary image pick-up card of a plurality of outputs; Wherein, select the far red light spectral coverage of whole spectrum medium wavelength between 9um and 11um as active far infrared light source, initiatively the far infrared light source is comprised of the N of wavelength between a 9.5um and 10.5um LED, and wherein N is natural number, N ∈ [26,36]; When the far infrared digital simulation field camera of the described active far infrared light source of combination and the subsidiary image pick-up card of described output, with the coaxial arrangement of far infrared digital simulation field camera of LED and the subsidiary image pick-up card of described output; The be del shape of described light emitting diode in the far infrared digital simulation field camera plane of the subsidiary image pick-up card of described output evenly put, with the light source stepless action in target image;
Collection for target image, use the target site in the visible light digital simulation field camera alignment image of the subsidiary image pick-up card of a described output to gather, and by image capture software, the far infrared digital simulation field camera of the visible light digital simulation field camera that is connected to the subsidiary image pick-up card of above-mentioned output on the same image pick-up card and the subsidiary image pick-up card of above-mentioned output is carried out synchronous acquisition and control, image acquisition has simultaneity;
The original image multilayer convergent-divergent of step 2, collection decomposes and shows; The original image convergent-divergent decomposition display packing of described collection is as follows: generate the target image thumbnail data according to the image acquisition data in step 1, the thumbnail data data type is the two-dimensional array that and display device show lattice match; Then generate corresponding thumbnail according to described target image thumbnail data, this is generated corresponding thumbnail, carry out turriform anisotropic filter group conversion more than 5 layers; Again the thumbnail of described bank of filters conversion rearranged the layering classification, thumbnail layer corresponding to del top in above-mentioned steps one is classified as the S1 layer, thumbnail layer corresponding to del bottom in step 1 is classified as the Sn layer, contain some layers between described S1 layer and Sn layer, n in its Sn is natural number, n ∈ [34,39]; Wherein, adjacent two thumbnail interlayer thumbnail image scale size ratios are 1～2: 1, and the thumbnail image of all thumbnail layers forms pyramid;
Step 3, target image identification; Described target image identification method is as follows: adopt linear SVM to classify to all the thumbnail layers in the pyramid in above-mentioned steps two, when classifying, choose 2/5ths to 3/5ths thumbnail layer as training set, then remaining thumbnail layer carries out the multiple dimensioned high-efficiency target image identification of multi-layer structure as test set as the standard of test accuracy by the equilibrium point on the demonstration curve that generates in described linear SVM.
2. the multi-scale self-adaptive high-efficiency target image identification method based on the multi-layer structure according to claim 1, it is characterized in that: the adjacent two thumbnail interlayer thumbnail image scale size ratios in above-mentioned steps two are 1.414: 1.
3. the multi-scale self-adaptive high-efficiency target image identification method based on the multi-layer structure according to claim 1, is characterized in that: contain 36 layers between the S1 layer in above-mentioned steps two and Sn layer.
4. the multi-scale self-adaptive high-efficiency target image identification method based on the multi-layer structure according to claim 1 is characterized in that: the thumbnail corresponding to this generation in above-mentioned steps two, carry out the conversion of 16 layers of turriform anisotropic filter group.
5. the multi-scale self-adaptive high-efficiency target image identification method based on the multi-layer structure according to claim 1, it is characterized in that: the thumbnail layer of choosing half in above-mentioned steps three is as training set.
6. the multi-scale self-adaptive high-efficiency target image identification method based on the multi-layer structure according to claim 1, it is characterized in that: the far red light spectral coverage that the whole spectrum medium wavelength of the selection in above-mentioned steps one is 10um is as active far infrared light source.
7. the multi-scale self-adaptive high-efficiency target image identification method based on the multi-layer structure according to claim 1, it is characterized in that: N the LED that the active far infrared light source in above-mentioned steps one is 10.5um by wavelength forms.
8. the multi-scale self-adaptive high-efficiency target image identification method based on the multi-layer structure according to claim 2, it is characterized in that: N the LED that the active far infrared light source in above-mentioned steps one is 10.5um by wavelength forms.
9. the multi-scale self-adaptive high-efficiency target image identification method based on the multi-layer structure according to claim 3, it is characterized in that: N the LED that the active far infrared light source in above-mentioned steps one is 10.5um by wavelength forms.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201110321561.8A CN102436576B (en) | 2011-10-21 | 2011-10-21 | Multi-scale self-adaptive high-efficiency target image identification method based on multi-level structure |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201110321561.8A CN102436576B (en) | 2011-10-21 | 2011-10-21 | Multi-scale self-adaptive high-efficiency target image identification method based on multi-level structure |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN102436576A CN102436576A (en) | 2012-05-02 |
| CN102436576B true CN102436576B (en) | 2013-11-06 |
Family
ID=45984631
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201110321561.8A Expired - Fee Related CN102436576B (en) | 2011-10-21 | 2011-10-21 | Multi-scale self-adaptive high-efficiency target image identification method based on multi-level structure |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN102436576B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107967454B (en) * | 2017-11-24 | 2021-10-15 | 武汉理工大学 | A two-way convolutional neural network remote sensing classification method considering the spatial neighborhood relationship |
Family Cites Families (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6389169B1 (en) * | 1998-06-08 | 2002-05-14 | Lawrence W. Stark | Intelligent systems and methods for processing image data based upon anticipated regions of visual interest |
| CN100568262C (en) * | 2007-12-29 | 2009-12-09 | 浙江工业大学 | Face recognition detection device based on multi-camera information fusion |
| CN201927079U (en) * | 2011-03-07 | 2011-08-10 | 山东电力研究院 | Rapid real-time integration processing system for visible image and infrared image |
-
2011
- 2011-10-21 CN CN201110321561.8A patent/CN102436576B/en not_active Expired - Fee Related
Also Published As
| Publication number | Publication date |
|---|---|
| CN102436576A (en) | 2012-05-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105224942B (en) | RGB-D image classification method and system | |
| Pflug et al. | Ear biometrics: a survey of detection, feature extraction and recognition methods | |
| Cao et al. | Marine animal classification using combined CNN and hand-designed image features | |
| Sugiarto et al. | Wood identification based on histogram of oriented gradient (HOG) feature and support vector machine (SVM) classifier | |
| Wang et al. | Object instance detection with pruned Alexnet and extended training data | |
| Zhang et al. | Distinguishing computer-generated images from natural images using channel and pixel correlation | |
| CN103646255A (en) | Face detection method based on Gabor characteristics and extreme learning machine | |
| Fiel et al. | Automated identification of tree species from images of the bark, leaves or needles | |
| Bouchaffra et al. | Structural hidden Markov models for biometrics: Fusion of face and fingerprint | |
| CN102637251A (en) | Face recognition method based on reference features | |
| CN105138998A (en) | Method and system for re-identifying pedestrian based on view angle self-adaptive subspace learning algorithm | |
| Hadid et al. | An Overview of Content-BasedImage Retrieval Methods And Techniques | |
| Siraj et al. | Flower image classification modeling using neural network | |
| Zhao et al. | Gabor-modulated grouped separable convolutional network for hyperspectral image classification | |
| Chen et al. | Image classification based on convolutional denoising sparse autoencoder | |
| CN102436576B (en) | Multi-scale self-adaptive high-efficiency target image identification method based on multi-level structure | |
| Shrivastava et al. | Bridging the semantic gap with human perception based features for scene categorization | |
| Yang et al. | Unsupervised domain adaptation for person re-identification based on multi-scale attention mechanism | |
| Jia et al. | Research of improved algorithm based on LBP for face recognition | |
| Yufeng et al. | Intelligent identification method of bicycle logo based on Haar classifier | |
| Huang et al. | Bark classification based on textural features using artificial neural networks | |
| Shariatmadar et al. | Sal-HMAX: An Enhanced HMAX Model in Conjunction With a Visual Attention Mechanism to Improve Object Recognition Task | |
| Cheng et al. | Retracted on March 1, 2022: A Review of Feature Extraction Technologies for Plankton Images | |
| Motamed et al. | Multimodal biometric recognition using particle swarm optimization-based selected features | |
| Ranjit et al. | A Symbolic Representation and Classification of Fruits. |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| DD01 | Delivery of document by public notice |
Addressee: Hong Tao Document name: Notification that Application Deemed not to be Proposed |
|
| GR01 | Patent grant | ||
| DD01 | Delivery of document by public notice |
Addressee: Hong Tao Document name: Notification to Pay the Fees |
|
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20131106 Termination date: 20141021 |
|
| EXPY | Termination of patent right or utility model | ||
| DD01 | Delivery of document by public notice |
Addressee: Hong Tao Document name: Notification of Decision on Request for Restoration of Right |