CN115641479B - Intelligent garbage classification change detection method based on attention mechanism - Google Patents
Intelligent garbage classification change detection method based on attention mechanism Download PDFInfo
- Publication number
- CN115641479B CN115641479B CN202211659770.8A CN202211659770A CN115641479B CN 115641479 B CN115641479 B CN 115641479B CN 202211659770 A CN202211659770 A CN 202211659770A CN 115641479 B CN115641479 B CN 115641479B
- Authority
- CN
- China
- Prior art keywords
- garbage
- image
- feature
- attention mechanism
- category
- 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.)
- Active
Links
- 239000010813 municipal solid waste Substances 0.000 title claims abstract description 102
- 230000008859 change Effects 0.000 title claims abstract description 40
- 238000001514 detection method Methods 0.000 title claims abstract description 34
- 230000007246 mechanism Effects 0.000 title claims abstract description 27
- 238000007781 pre-processing Methods 0.000 claims abstract description 8
- 238000012549 training Methods 0.000 claims description 35
- 238000000034 method Methods 0.000 claims description 21
- 238000012545 processing Methods 0.000 claims description 12
- 239000013598 vector Substances 0.000 claims description 10
- 230000004913 activation Effects 0.000 claims description 7
- 238000004364 calculation method Methods 0.000 claims description 7
- 238000013528 artificial neural network Methods 0.000 claims description 6
- 230000004069 differentiation Effects 0.000 claims description 5
- 230000006870 function Effects 0.000 claims description 4
- 230000003993 interaction Effects 0.000 claims description 4
- 239000010806 kitchen waste Substances 0.000 claims description 3
- 238000011176 pooling Methods 0.000 claims description 3
- 230000002452 interceptive effect Effects 0.000 claims description 2
- 238000005259 measurement Methods 0.000 claims description 2
- 230000035945 sensitivity Effects 0.000 claims description 2
- 230000000873 masking effect Effects 0.000 claims 1
- 230000004660 morphological change Effects 0.000 abstract description 2
- 230000017105 transposition Effects 0.000 abstract description 2
- 238000013527 convolutional neural network Methods 0.000 description 9
- 238000010586 diagram Methods 0.000 description 8
- 230000011218 segmentation Effects 0.000 description 7
- 238000013461 design Methods 0.000 description 6
- 238000005516 engineering process Methods 0.000 description 3
- 238000003709 image segmentation Methods 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000002372 labelling Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 230000008520 organization Effects 0.000 description 2
- 230000003213 activating effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 238000005286 illumination Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000005728 strengthening Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02W—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO WASTEWATER TREATMENT OR WASTE MANAGEMENT
- Y02W30/00—Technologies for solid waste management
- Y02W30/10—Waste collection, transportation, transfer or storage, e.g. segregated refuse collecting, electric or hybrid propulsion
Landscapes
- Image Analysis (AREA)
Abstract
The invention provides an intelligent garbage classification change detection method based on an attention mechanism, which comprises the steps of continuously acquiring target images within a set range of image equipment; deploying Slim Swin Transformer network structure to construct AI reasoning model; and preprocessing an input target image based on the AI reasoning model to obtain a garbage area and a corresponding garbage category generated by the current delivery. The AI model provided by the invention encodes through a multi-head attention mechanism, obtains transposition bias of local correlation and translational invariance through a sliding window mechanism, and simultaneously enables the model to understand garbage images on high-level semantics after semantic understanding capability and long-distance modeling capability of the attention mechanism, so that the movement, morphological change and overlapping coverage position relation of objects are better understood, and the AI model is different from the prior art, and better handles noise interference.
Description
Technical Field
The invention relates to the technical field of garbage classification monitoring, in particular to an intelligent garbage classification change detection method based on an attention mechanism.
Background
In the field of images, the recognition, segmentation and classification of images by constructing a neural network model through a deep learning technology are the most advanced and optimal technical scheme.
The most important technology used for detecting garbage change areas and classification is change detection in intelligent image vision and image segmentation.
In the specific garbage change detection and classification, in the existing scheme:
(1) the scheme using the multi-layer convolutional neural network model has poor segmentation capability, because the convolutional neural network is not good at constructing the relation of pixel blocks, the granularity of the segmentation is rough, and large-scale connection exists between garbage. And under the condition of higher segmentation semantic categories, the multi-layer convolutional neural network is difficult to encode complex different semantics, so that the multi-layer convolutional neural network is not easy to adapt to multi-category image segmentation. The accuracy of this scheme is poor.
(2) Meanwhile, if the traditional algorithm is used, the calculation complexity is high, so that the consumption of calculation resources is high, and the scheme lacks variability due to the utilization of a fixed filter kernel and a database, so that the limitation of required conditions is excessive, and the robustness is lacking. Edges in the garbage image have the characteristic of sharpness, and the color change is various and ambiguous, so that false detection can be generated by utilizing filtering detection. The fixability of the database cannot be well adapted to the form and color change of garbage when the color of the image is shifted due to illumination and the like, so that the scheme is missed.
(3) Finally, if a manual detection scheme is used, the cost is too high, and the detection throughput is obviously low. The scheme of adopting manual detection generally limits the user to throw garbage in fixed time, and causes obvious inconvenience. And because this solution is based on personnel labor, such a rubbish detection system has a high economic cost.
Based on the above, the invention is urgently needed to design an intelligent garbage classification change detection method which can obtain local correlation through a sliding window mechanism and has semantic understanding capability and long-distance modeling capability of an attention mechanism so as to better cope with noise interference.
Disclosure of Invention
Aiming at the defects existing in the prior art, the invention aims to provide the intelligent garbage classification change detection method based on the attention mechanism, and the invention has the advantage of being more convenient to update and deploy compared with the manual design characteristic relation and the search analysis of a database of the traditional image processing scheme by proposing a mode of training and reasoning in an end-to-end mode. To solve the problems set forth in the background art.
In order to achieve the above object, the present invention is realized by the following technical scheme: intelligent garbage classification change detection method based on attention mechanism comprises the following steps
Continuously acquiring a target image within a set range of image equipment, and uploading the target image to a server side;
deploying Slim Swin Transformer network structure to construct AI reasoning model to reduce computing resource of server;
preprocessing an input target image based on an AI reasoning model to obtain a garbage area generated by current delivery and a corresponding garbage category;
analyzing the garbage area and the garbage category to generate a predicted result value for garbage classification, uploading the predicted result value to a server, and judging whether illegal delivery is performed when a user delivers garbage the current time based on the predicted result value: if yes, submitting the target image to manual accounting processing and then returning to acquire the target image again; and otherwise, ending the detection.
As an improvement of the invention, the method for obtaining the garbage area and the corresponding garbage category result generated by the current delivery comprises the following steps: the method for obtaining the garbage area and the corresponding garbage category result generated by the current delivery comprises the following steps:
s1, acquiring a data set of the AI reasoning model to train the AI reasoning model
S1-1, acquiring a same dustbin at the current time based on image equipment, and acquiring a front photographed image and a rear photographed image which are continuous in time sequence;
s1-2, pairing of a change area of the acquired rear shooting image relative to the front shooting image and a garbage category to which the change area belongs is realized through manual marking;
s1-3, repeating the steps S1-1 and S1-2 until the paired shooting images marked manually reach a set quantity, and taking the paired shooting images as an acquisition training sample, wherein at least first training characteristic information and second training characteristic information are marked in the training sample, the first training characteristic information is used for representing a change region of the shooting images, and the second training characteristic information is used for representing a garbage category to which the change region belongs;
s1-4, inputting the first training feature information and the second training feature information into Slim Swin Transforme for training at the same time, and performing differentiation processing on the features at each stage by utilizing the mixed Stages to obtain an AI reasoning model;
s2, after normalizing and size preprocessing is carried out on the target image, inputting the processed target image into an AI reasoning model
S2-1, an AI reasoning model encodes an input target image into an embedded vector through a convolution layer, the embedded vector is sent to a transform layer, and a multi-level feature map is constructed in different modes according to a token mixing mode so as to enable a network to adapt to targets with different scales;
s2-2, sending feature graphs obtained by different levels into a token mixer, wherein a movable window is adopted in an attention module to perform inter-window feature interaction so as to extract richer features;
s2-3, inputting the feature map into a difference detection module constructed based on a neural network after splicing the feature map, performing secondary splicing after obtaining a difference feature map used for representing the feature map, and performing convolution and activation function pretreatment on the difference feature map subjected to secondary splicing based on a classification module to obtain a garbage area and a corresponding garbage category result generated by current delivery.
Compared with the prior art, the invention has the beneficial effects that:
1. the AI model provided by the invention encodes through a multi-head attention mechanism, obtains transposition bias of local correlation and translational invariance through a sliding window mechanism, and simultaneously enables the model to understand garbage images on high-level semantics after semantic understanding capability and long-distance modeling capability of the attention mechanism, so that the movement, morphological change and overlapping coverage position relation of objects are better understood, and the AI model is different from the prior art, and better handles noise interference;
2. the AI model provided by the invention is trained and inferred in an end-to-end manner, and compared with the manual design characteristic relation of the traditional image processing scheme and the search analysis of a database, the AI model is more convenient to update and deploy;
3. the image data reported by each garbage delivery point is intensively processed in the background of the server by the end of the server model, so that the labor cost is obviously reduced, and the popularization and implementation of the garbage classification detection system in each region are facilitated.
Drawings
The disclosure of the present invention is described with reference to the accompanying drawings. It should be understood that the drawings are for purposes of illustration only and are not intended to limit the scope of the present invention in which like reference numerals are used to designate like parts. Wherein:
FIG. 1 is a schematic diagram of an AI reasoning model in accordance with one embodiment of the invention;
FIG. 2 is a schematic diagram of a Slim Swin Transformer network architecture designed based on a MetaFormer architecture according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of Stage1 and Stage4 using a conventional attention mechanism as a token mixer when differentiating features according to an embodiment of the present invention;
FIG. 4 is a schematic diagram of the pooling used as Stage2 and Stage3 token mixer in differentiating features according to one embodiment of the present invention;
FIG. 5 is a schematic diagram of determining whether to deliver a violation when delivering a rubbish from a delivering user according to an embodiment of the present invention;
FIG. 6 is a schematic diagram illustrating training and deployment of AI reasoning models in a spam delivery system in accordance with one embodiment of the present invention;
fig. 7 is a schematic diagram of preprocessing an input target image by an AI inference model according to an embodiment of the present invention to obtain a classification result.
Detailed Description
It is to be understood that, according to the technical solution of the present invention, those skilled in the art may propose various alternative structural modes and implementation modes without changing the true spirit of the present invention. Accordingly, the following detailed description and drawings are merely illustrative of the invention and are not intended to be exhaustive or to limit the invention to the precise form disclosed.
The present invention will be described in further detail with reference to the accompanying drawings, which are not intended to limit the invention.
As the technical conception and the realization principle of the invention are understood, the scheme using the multi-layer convolutional neural network model has poor segmentation capability, and because the convolutional neural network is not good at constructing the relation of pixel blocks, the granularity of the segmentation is rough, and large-scale connection exists between garbage. And under the condition of higher segmentation semantic categories, the multi-layer convolutional neural network is difficult to encode complex different semantics, so that the multi-layer convolutional neural network is not easy to adapt to multi-category image segmentation. The accuracy of the prior art solutions is thus poor.
Therefore, compared with a model based on a multi-layer convolution neural network, the AI reasoning model constructed by deploying the Slim Swin Transformer network structure is provided by the invention, the existing attention mechanism is obviously better than the coding capacity and long-distance modeling capacity of the convolution neural network, the generalized bias of the local correlation and translational invariance of the convolution neural network is also provided, the garbage shape and position change can be well adapted, the semantic relation between garbage targets is understood, the covering and impact movement among objects are well robust, the false detection is obviously reduced,
but to realize the technical conception and solve the defects of the prior technical proposal.
As shown in fig. 1, as an embodiment of the present invention, an intelligent garbage classification change detection method based on an attention mechanism is proposed, including:
the first step, continuously obtaining target images within a set range of the image equipment, and uploading the target images to a server side.
Based on the above technical concept, it can be understood that, in a specific implementation, a worker can shoot and acquire an image from a camera shooting installed below a cover of the dustbin or a device inside the dustbin in a overlooking manner, upload the image to a server, generate an image each time a user puts in garbage, and pair shooting images, wherein each pair of images is two consecutive shooting images (A, B) in time sequence for the same dustbin.
And secondly, deploying Slim Swin Transformer network structure to construct an AI reasoning model so as to reduce the computing resources of the server side.
It should be noted that, slim Swin Transformer is inspired by the design of MetaFormer about the converter architecture, and the design of the converter architecture rather than the research result of the token mixing method is performed, so that the invention adopts Hybrid Stages to perform differentiation processing on the multi-level features, and the method of using token mixer in the architecture is mixed, and the main architecture is shown in fig. 2.
Thirdly, preprocessing an input target image based on an AI reasoning model to obtain a garbage area and a corresponding garbage category generated by the current delivery.
As shown in fig. 6 to 7, based on the above technical concept, it should be noted that the method for obtaining the garbage area and the corresponding garbage category result generated by the current delivery includes:
s1, acquiring a data set of an AI (advanced technology attachment) reasoning model to train the AI reasoning model
S1-1, acquiring a same dustbin at the current time based on image equipment, and acquiring a front photographed image and a rear photographed image which are continuous in time sequence;
s1-2, pairing of the change area of the acquired rear shooting image relative to the front shooting image and the garbage category to which the change area belongs is achieved through manual labeling, and the method is as follows: each pair of shot images are two continuous shot images (A, B) of the same garbage can in time sequence, the organization of the data set is realized by manual labeling, and the change area of the B relative to the A and the class of garbage to which the change area belongs are labeled;
s1-3, repeating the steps S1-1 and S1-2 until the paired shooting images marked manually reach a set quantity, and taking the paired shooting images as an acquisition training sample, wherein at least first training characteristic information and second training characteristic information are marked in the training sample, the first training characteristic information is used for representing a change area of the shooting images, and the second training characteristic information is used for representing garbage category to which the change area belongs;
after the AI reasoning model is deployed, a user shoots and sends an image in a overlooking mode through card swiping delivery each time, after receiving a photo B returned by a specific garbage can, a server inputs the photo B and a corresponding time sequence previous photo A into the AI model together, and at the moment, the following steps are implemented:
s1-4, inputting first training feature information and second training feature information into Slim Swin Transforme for training at the same time, and performing differentiation processing on the features at each stage by using the mixed Stages to obtain an AI reasoning model; it can be understood that the AI reasoning model segments the meaningful change region in the photo B relative to the photo a by reasoning, and marks the garbage category to which the segmented region belongs, and the specific method for differentiating the features in the specific implementation comprises: stage1 and Stage4 use the traditional attention mechanism as a token mixer, as shown in fig. 3, to extract more abundant features in Stage1, and at the same time, integrate the features extracted by the previous three stages better in Stage 4. Since there is no weight parameter to be learned in the manner of using pooling as Stage2 and Stage3 token mixer, as shown in fig. 4, this step makes each token average to aggregate the features of its nearby token, and finally, the invention also proves that the above-mentioned mixer method employed in differentiating the features maintains high accuracy while also reducing the parameter amount of the model;
s2, after normalization and size preprocessing are carried out on the target image, inputting the processed target image into an AI reasoning model, wherein the AI reasoning model specifically comprises the following steps:
s2-1, an AI reasoning model encodes an input target image into an embedded vector through a convolution layer, the embedded vector is sent to a transform layer, a multi-level feature map is constructed in different modes according to a token multiplexing mode, a network is adapted to targets with different scales, the attention mechanism of the image transform layer is to make local self-attention on the feature map through a sliding window, the embedded vector is mapped into three vectors of q, k and v, then for each embedded vector, matrix multiplication operation is respectively carried out on k corresponding to all the vectors of q, attention weight values are obtained after activating the function, all v are multiplied by the weight values, attention output values are obtained, and in specific implementation, the multi-level feature map is constructed, and the specific implementation of interactive window self-attention calculation comprises the following steps: referring to the design concept of the convolutional neural network level, the Patch metering modules are utilized in Stage1 and Stage4 to select elements at intervals in the row direction and the column direction for splicing, then the full-connection layer is utilized for carrying out Concat, the Patch metering modules in Stage2 and Stage3 parts are utilized for carrying out downsampling before each Stage is started, and based on the fact, multiple levels are constructed through the steps, and feature diagrams with different scales are obtained;
s2-2, sending feature graphs obtained by different levels to a token mixer, wherein the attention module adopts a moving window to perform inter-window feature interaction so as to extract richer features, and the specific method of the attention module adopting the moving window to perform interaction comprises the following steps: the windows are moved downwards and leftwards, the number of the shifted latches is kept unchanged through the shift splicing of part of the windows, and the purpose of the method is to increase the receptive field of each window without increasing the calculated amount. Adding a mask operation to windows which are not associated before shifting but are adjacent after shifting, wherein the mask operation is used for ensuring that non-adjacent modules do not perform attention calculation;
s2-3, inputting the spliced feature images into a difference detection module constructed based on a neural network, performing secondary splicing after obtaining the difference feature images used for representing the feature images, and performing convolution and activation function pretreatment on the difference feature images subjected to secondary splicing based on a classification module to obtain garbage areas and corresponding garbage category results generated by current delivery, wherein the specific implementation needs to be noted that the specific method for performing convolution activation treatment after performing secondary splicing on the difference feature images based on the difference detection module comprises the following steps: receiving characteristics of different scales generated by different levels of a network, calculating differences of the front and rear images, (distinguishing a mode of absolute difference calculation in a previous BIT or directly) Slim Swin Transformer, splicing the change images of each level, performing convolution, activation and BN operation on the spliced difference characteristics, and further extracting the difference characteristics, wherein the aim of doing so is to learn the optimal distance measurement of each scale during training and improve sensitivity to a change region as shown in fig. 4;
fourthly, analyzing the garbage area and garbage category to generate a predicted result value for garbage classification, uploading the predicted result value to a server, and judging whether illegal delivery is performed when a user delivers garbage at the present time based on the predicted result value: if yes, submitting the target image to manual accounting processing and then returning to acquire the target image again; and otherwise, ending detection, wherein the specific method for analyzing the garbage area and generating the predicted result value for garbage classification by the garbage classification comprises the following steps of: as shown in fig. 5, the features output by the four level difference modules of different levels are aligned in feature scale through MLP, the classification result of each pixel on the current feature scale is output through a convolution classifier, and finally the result on the current feature scale is required to be restored to the original image scale; the specific method for judging whether illegal delivery occurs when the user delivers the garbage at the present time comprises the following steps: the garbage can is divided into kitchen waste, harmful, recyclable and other four major categories, the garbage sub-category index contained in each garbage can is established in advance, such as that a battery belongs to harmful garbage, a plastic bottle belongs to recyclable, at this time, it can be understood that the network output is that the later image belongs to the category compared with the newly-added area of the former image, then a user only needs to judge whether the category belongs to the garbage can or not, the illegal placement is not performed, and the garbage bags placed in the kitchen waste garbage can are illegal garbage.
It can be understood that after the first to fourth steps are completed, the end-to-end mode is completed for training and reasoning, at this time, the end of the model is located at the server, and the image data reported by each garbage delivery point is centrally processed at the server background, so that the labor cost is obviously reduced, and the popularization and implementation of the garbage classification detection system in each region are facilitated.
However, in order to ensure the correctness of the centralized processing of the image data reported by each garbage delivery point in the background of the server, the AI inference model needs to be optimized, that is, the trained AI inference model needs to be verified in recall and accuracy to optimize the AI inference model.
In an embodiment of the present invention, the AI reasoning model code provided by the present invention may be implemented by a Pytorch framework, and run on a server of the Linux system, which is equipped with the NVIDIA Tesla V100 SXM2 32 GB graphics card. Training the model through a data set of 10000 pictures of an enterprise organization, and using a color gamut conversion and random clipping strengthening scheme, training a set and a test set to 9:1 division. After the classification index converges to the minimum value, the classification index is deployed on a server, an inference interface is provided, and an input image is inferred through a Pytorch framework input model. According to the existing experimental results, the scheme has better performance in training time, reasoning speed and space occupation, and can achieve the effect of detecting the expected and excellent illegal delivery of the garbage.
The technical scope of the present invention is not limited to the above description, and those skilled in the art may make various changes and modifications to the above-described embodiments without departing from the technical spirit of the present invention, and these changes and modifications should be included in the scope of the present invention.
Claims (7)
1. The intelligent garbage classification change detection method based on the attention mechanism is characterized by comprising the following steps of: comprising the following steps:
continuously acquiring a target image within a set range of image equipment, and uploading the target image to a server side;
deploying Slim Swin Transformer network structure to construct AI reasoning model to reduce computing resource of server;
preprocessing an input target image based on an AI reasoning model to obtain a garbage area generated by current delivery and a corresponding garbage category; wherein,,
the method for obtaining the garbage area and the corresponding garbage category result generated by the current delivery comprises the following steps:
s1, acquiring a data set of the AI reasoning model to train the AI reasoning model
S1-1, acquiring a same dustbin at the current time based on image equipment, and acquiring a front photographed image and a rear photographed image which are continuous in time sequence;
s1-2, pairing of a change area of the acquired rear shooting image relative to the front shooting image and a garbage category to which the change area belongs is realized through manual marking;
s1-3, repeating the steps S1-1 and S1-2 until the paired shooting images marked manually reach a set quantity, and taking the paired shooting images as an acquisition training sample, wherein at least first training characteristic information and second training characteristic information are marked in the training sample, the first training characteristic information is used for representing a change region of the shooting images, and the second training characteristic information is used for representing a garbage category to which the change region belongs;
s1-4, inputting the first training feature information and the second training feature information into Slim Swin Transforme for training at the same time, and performing differentiation processing on the features at each stage by utilizing the mixed Stages to obtain an AI reasoning model;
the specific method for carrying out differentiation processing on the characteristics comprises the following steps:
stage1 and Stage4 use the traditional attention mechanism as a token mixer to extract richer features in Stage1, so that the features extracted by the previous three stages are integrated in Stage 4;
the pooling is adopted as Stage2 and Stage3 token mixer, and is used for enabling each token to average and aggregate the characteristics of nearby tokens, so that the parameter quantity of the model is reduced while high precision is maintained;
s2, after normalizing and size preprocessing is carried out on the target image, inputting the processed target image into an AI reasoning model
S2-1, an AI reasoning model encodes an input target image into an embedded vector through a convolution layer, the embedded vector is sent to a transform layer, and a multi-level feature map is constructed in different modes according to a token mixing mode so as to enable a network to adapt to targets with different scales;
s2-2, sending feature graphs obtained by different levels into a token mixer, wherein a movable window is adopted in an attention module to perform inter-window feature interaction so as to extract richer features;
s2-3, inputting the feature images into a difference detection module constructed based on a neural network after splicing the feature images, performing secondary splicing after obtaining a difference feature image used for representing the feature images, and performing convolution and activation function pretreatment on the difference feature images subjected to secondary splicing based on a classification module to obtain a garbage area generated by current delivery and a corresponding garbage category result;
analyzing the garbage area and the garbage category to generate a predicted result value for garbage classification, uploading the predicted result value to a server, and judging whether illegal delivery is performed when a user delivers garbage the current time based on the predicted result value: if yes, submitting the target image to manual accounting processing and then returning to acquire the target image again; and otherwise, ending the detection.
2. The attention mechanism-based intelligent garbage classification change detection method as claimed in claim 1, wherein: the Slim Swin Transformer network structure is designed based on a MetaFormer architecture.
3. The attention mechanism-based intelligent garbage classification change detection method as claimed in claim 1, wherein: in step S2-1, a multi-level feature map is constructed, and the specific method for performing interactive window self-attention calculation comprises the following steps:
the Patch merge module is utilized in Stage1 and Stage4 to select elements at intervals in the row and column directions for splicing
Then using the full connection layer to perform Concat;
the Patch Embedding modules of Stage2 and Stage3 part utilize convolution stride to downsample before each Stage starts, and construct a multi-level feature map so as to obtain feature maps of different scales.
4. The attention mechanism-based intelligent garbage classification change detection method as claimed in claim 1, wherein: in step S2-2, the specific method for the attention module to interact by adopting the moving window comprises the following steps:
the window is moved downwards and leftwards, the number of the latches after the movement is kept unchanged through the shift splicing of part of the windows, and the receptive field of each window is increased while the calculated amount is not increased;
masking operations are added to windows that are not associated before the shift but are adjacent after the shift to ensure that non-adjacent modules do not perform attention calculations.
5. The attention mechanism-based intelligent garbage classification change detection method as claimed in claim 1, wherein: in step S2-3, the specific method for performing convolution activation processing after performing secondary splicing on the difference feature map based on the difference detection module comprises the following steps:
receiving characteristics of different scales generated based on different levels of a network, and calculating differences of the front image and the rear image;
slim Swin Transformer splicing the change graphs of each level, and performing convolution, activation and BN operation on the spliced difference features;
the difference features are further extracted to achieve the best distance measurement for learning each scale during training, and sensitivity to the change area is improved.
6. The attention mechanism-based intelligent garbage classification change detection method as claimed in claim 1, wherein: the specific method for analyzing the garbage area and generating the predicted result value for garbage classification by the garbage classification comprises the following steps: and (3) carrying out feature scale alignment on the features output by the different-level difference detection modules through the MLP, outputting each pixel classification result on the current feature scale through the convolution classifier, and finally restoring the result on the current feature scale to the original image scale to generate a predicted result value of garbage classification.
7. A method for detecting intelligent garbage classification change based on an attention mechanism according to claim 1 or 3, characterized in that: the specific method for judging whether illegal delivery is carried out when the user delivers the garbage at the current time comprises the following steps:
the garbage can is divided into kitchen waste, harmful, recyclable and other four categories;
establishing a garbage sub-class index contained in each garbage bin in advance;
and obtaining the category of the newly added area of the next image based on network output compared with the category of the newly added area of the previous image, and judging whether the category belongs to the garbage can or not, wherein the condition that the category does not belong to the garbage can is illegal delivery.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211659770.8A CN115641479B (en) | 2022-12-23 | 2022-12-23 | Intelligent garbage classification change detection method based on attention mechanism |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211659770.8A CN115641479B (en) | 2022-12-23 | 2022-12-23 | Intelligent garbage classification change detection method based on attention mechanism |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115641479A CN115641479A (en) | 2023-01-24 |
| CN115641479B true CN115641479B (en) | 2023-08-15 |
Family
ID=84949638
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211659770.8A Active CN115641479B (en) | 2022-12-23 | 2022-12-23 | Intelligent garbage classification change detection method based on attention mechanism |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115641479B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN119250373A (en) * | 2024-12-04 | 2025-01-03 | 上海环联生态科技有限公司 | A garbage classification evaluation system |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113255804A (en) * | 2021-06-03 | 2021-08-13 | 图灵人工智能研究院（南京）有限公司 | Garbage traceability method and device based on image change detection |
| CN113657143A (en) * | 2021-06-25 | 2021-11-16 | 中国计量大学 | A Garbage Classification Method Based on Joint Decision of Classification and Detection |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7773813B2 (en) * | 2005-10-31 | 2010-08-10 | Microsoft Corporation | Capture-intention detection for video content analysis |
-
2022
- 2022-12-23 CN CN202211659770.8A patent/CN115641479B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113255804A (en) * | 2021-06-03 | 2021-08-13 | 图灵人工智能研究院（南京）有限公司 | Garbage traceability method and device based on image change detection |
| CN113657143A (en) * | 2021-06-25 | 2021-11-16 | 中国计量大学 | A Garbage Classification Method Based on Joint Decision of Classification and Detection |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115641479A (en) | 2023-01-24 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Yan et al. | Automated asphalt highway pavement crack detection based on deformable single shot multi-box detector under a complex environment | |
| CN114241053B (en) | Multi-category tracking method based on improved attention mechanism FairMOT | |
| Golovko et al. | Development of solar panels detector | |
| CN105574550A (en) | Vehicle identification method and device | |
| CN111368690A (en) | Deep learning-based video image ship detection method and system under influence of sea waves | |
| CN102915432B (en) | A kind of vehicle-mounted microcomputer image/video data extraction method and device | |
| CN109993138A (en) | A kind of car plate detection and recognition methods and device | |
| CN103714321A (en) | Driver face locating system based on distance image and strength image | |
| CN109886147A (en) | A vehicle multi-attribute detection method based on single-network multi-task learning | |
| CN115641479B (en) | Intelligent garbage classification change detection method based on attention mechanism | |
| Baoyuan et al. | Research on object detection method based on FF-YOLO for complex scenes | |
| Cao et al. | An end-to-end neural network for multi-line license plate recognition | |
| Pan et al. | A hybrid deep learning algorithm for the license plate detection and recognition in vehicle-to-vehicle communications | |
| CN114973067A (en) | Earth pressure balance shield muck information extraction method based on deep multi-task learning | |
| Elihos et al. | Deep learning based segmentation free license plate recognition using roadway surveillance camera images | |
| Haq et al. | Rotation aware object detection model with applications to weapons spotting in surveillance videos | |
| CN113191597A (en) | Commonweal litigation evidence obtaining method and device based on knowledge graph and unmanned aerial vehicle | |
| CN117541824A (en) | Identification method and device, storage medium and electronic equipment | |
| CN115359487B (en) | A rapid rail vehicle number recognition method, device and storage medium | |
| Lu et al. | Detection of small objects in complex long-distance scenes based on Yolov3 | |
| Huang et al. | Key technologies of intelligent transportation based on image recognition and optimization control | |
| CN118552792B (en) | Device detection method, system, computer storage medium and electronic device | |
| Jagtap et al. | Suspicious vehicle recognition using number plate | |
| Zhao et al. | A novel deep learning framework for recognizing high-speed train numbers | |
| CN120144784B (en) | Large-model-based data service system of combined combat target system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |