Target detection method and system based on multi-scale tokens
By introducing multi-scale internal token modules and differential filter residual modules into the target detection network, focusing on the multi-scale characteristics and local spatial characteristics within a single tag, the problem of semantic information loss in the prior art is solved, and more optimized object detection performance is achieved.
Patent Information
- Application Number
- CN202211463328.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-11-22
- Publication Date
- 2025-04-22
- Estimated Expiration
- 2042-11-22
AI Technical Summary
When focusing on the multi-scale characteristics between tokens and tokens, existing object detection technologies ignore fine-grained features within a single tag, resulting in the loss of semantic information and affect detection performance.
Using a multi-scale token-based object detection method, the object detection network is built, including multiple multi-scale internal token modules and differential filter residual modules of different scales, focusing on the multi-scale characteristics and local spatial characteristics within a single input mark, and enhancing fine-grained feature extraction and edge information learning.
It effectively avoids the loss of semantic information, optimizes the target detection performance, and can achieve the most advanced results in the target detection task.
Smart Images
Figure CN116051882B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of computer vision and relates to a target detection method and system based on multi-scale tokens. Background Art
[0002] Object detection has always been a core task in the field of computer vision. Computers collect, store and learn real-world images, extract deep features, and ultimately accurately and efficiently capture the area of interest in the image, draw a bounding box around the target, and obtain its category information and two-dimensional coordinate information. With the development of the intelligent and information age, object detection technology is increasingly penetrating into practical applications, such as autonomous driving, face recognition, public safety, etc., and has great practical research significance and value in academia or business.
[0003] Currently, visual transformers have achieved impressive performance in object detection. Traditional object detection only focuses on the multi-scale characteristics between tokens and tokens, but does not pay attention to the fine-grained features within a single token, which may lead to the loss of semantic information in the object detection task. Summary of the invention
[0004] The purpose of the present invention is to provide a target detection method and system based on multi-scale tokens to avoid semantic information loss and optimize target detection performance.
[0005] In order to achieve the above object, the basic scheme of the present invention is: a target detection method based on multi-scale tokens, comprising the following steps:
[0006] Get the original image and convolve it into image blocks of two sizes, a and b;
[0007] Constructing a target detection network, the target detection network includes a plurality of multi-scale internal token modules of different scales and a plurality of corresponding difference filter residual modules;
[0008] The image patch is input into the multi-scale internal token module and the difference filter residual module of the first stage;
[0009] The multi-scale output of the multi-scale internal token module in the previous stage and the output of the difference filter residual module in the previous stage are used as the input of the multi-scale internal token module in the next stage, and the output of the difference filter residual module in the previous stage is used as the input of the difference filter residual module in the next stage;
[0010] The output of the multi-scale internal token module in the final stage is fused with the output of the difference filter residual module in the final stage to obtain the target detection result.
[0011] The working principle and beneficial effects of this basic scheme are: multi-scale internal token module and differential filter residual module. The multi-scale internal token module can enrich the fine-grained features by focusing on the multi-scale features within a single input token. The differential filter residual module, as a convolutional residual network, helps to enhance the local spatial features of a single token and learn more edge information. This scheme can achieve state-of-the-art performance in object detection.
[0012] Furthermore, the method for obtaining multi-scale output by the multi-scale internal token module is as follows:
[0013] The image tokens are mapped to three matrices Q, K, and V. The matrices K and V are downsampled to different scales using convolution:
[0014]
[0015] K i,j =MS j (K i ), j=1,2,3...
[0016] V i,j =MS j (V i ), j = 1, 2, 3...
[0017]
[0018] Among them, MS is a multi-scale function, i represents the sampling stage, j represents different scales; Q is the query matrix, K represents the matrix of the correlation between the queried information and other information, V represents the matrix of the queried information; X is the feature map, W is the weight matrix; W Qi , W Ki , W Vi Indicates that a corresponding weight matrix is assigned to the feature map X; Softmax is a normalization function, Attention is the symbolic mark of self-attention; T represents transposition, d k Represents the dimension of matrix K; MS j (K i ) indicates that the matrix K is transformed to the jth scale in the i-th stage, MS j (V i ) means that in the i-th stage, the matrix V is transformed to the j-scale size, Q i , K i , V i The query matrix, the correlation matrix, and the matrix of the queried information calculated in the i-th sampling stage;
[0019] Through multi-scale transformation, we can get matrices K of different sizes. i,j and V i,j , by upsampling the matrix Ki,j and V i,j The matrix sizes are unified and then spliced;
[0020] For the matrix K i,j and V i,j Rearrange the elements in rows and columns with a preset interval, K j and V j Form new key-value pairs, expand the receptive domain, and achieve global perception without losing information;
[0021] A nonlinear transformation layer is applied to expand the search feature subspace. The nonlinear transformation layer includes a linear layer, a relu layer, a pooling layer, and a linear layer connected in sequence to fit richer image features. Combined with the linear transformation layer, more combinations of Q, K, and V are found to mine the fine-grained features of the mark. The linear transformation layer uses a fully connected layer to connect the transformation. The nonlinear transformation layer and the linear transformation layer are composed of two fully connected layers, a pooling layer, and an activation function:
[0022] Q non =f(Activate(f(Q))+b)
[0023] K non =f(Activate(f(K))+b)
[0024] V non =f(Activate(f(V))+b)
[0025] Among them, f is a linear function, Activate is an activation function, and b is a bias;
[0026] Using the dual feature extraction method, the original image is convolved with different convolution kernels to obtain different feature maps to enrich the spatial context and fine-grained features of the input tag. For the input original image X, a non-overlapping sliding window of size 1*1 is used to obtain the embedded tag, focusing on the fine-grained features; an overlapping sliding window of size 3*3 and a preset stride is used to capture the spatial context.
[0027] The convolution operation can learn local context to enrich semantic features, making up for the defect that the model can only capture global context information. Focusing attention on different levels is conducive to the learning of potential fine-grained features within a single token.
[0028] Furthermore, the method for extracting high-frequency information of the input end by the difference filter residual module is as follows:
[0029] Assume the size of the image token is C*H*w, apply average pooling to obtain smooth features, and then use two convolution kernels to obtain two different smooth images GA , G B :
[0030] G A ,G B = Cov(Avgpool(G))
[0031] Among them, Cov represents the convolution function, the convolution kernel size is 3*3, the stride is 1, and the padding is 1; Avgpool represents the pooling function, the average pool kernel size is 2*2, and the stride is 2; G is the original graph;
[0032] G A , G B Subtract the difference image, and then upsample it to restore the size of the feature map.
[0033] The difference filter residual module is used to extract high-frequency information from the input and reduce semantic ambiguity in the attention mechanism.
[0034] Further, the method to obtain the target detection result is as follows:
[0035] Calculate and fuse based on the semantic importance of the edge information of the image:
[0036]
[0037]
[0038] Among them, f is a linear function that is similar to a 1*1 convolutional layer; T is the input feature map matrix, H and w are the height and width of the feature map, Gaus represents the target detection function, ζ is the normalization function, and x is the variable of the function, that is, the matrix pixel value.
[0039] The calculation is simple and easy to operate.
[0040] The present invention also provides a target detection system based on multi-scale tokens, including a data acquisition module and a processing module. The data acquisition module is used to acquire the original image. The output end of the data acquisition module is connected to the input end of the processing module. The processing module executes the method of the present invention to perform target detection.
[0041] The system can focus on the internal features of each token and effectively extract more multi-scale features before the self-attention layer. It also helps the transformer to better learn edge features and achieve state-of-the-art performance.
[0042] Further, the processing module includes a plurality of multi-scale internal token modules of different scales and a plurality of differential filter residual modules corresponding thereto;
[0043] The output end of the data acquisition module is respectively connected to the input end of the multi-scale internal token module of the first stage and the input end of the difference filter residual module of the first stage;
[0044] The multi-scale output of the multi-scale internal token module in the previous stage and the output of the difference filter residual module in the previous stage are used as the input of the multi-scale internal token module in the next stage, and the output of the difference filter residual module in the previous stage is used as the input of the difference filter residual module in the next stage;
[0045] The output of the multi-scale internal token module in the final stage is fused with the output of the difference filter residual module in the final stage to output the target detection result.
[0046] A multi-scale internal token module and multiple difference filter residual modules are set up to optimize target detection and facilitate use.
[0047] Furthermore, the multi-scale internal token module includes a multi-scale operation unit, a splicing unit, a linear layer and a non-linear layer, the output end of the multi-scale operation unit is connected to the input end of the splicing unit, and the output end of the splicing unit is connected to the input ends of the linear layer and the non-linear layer respectively.
[0048] The structure is simple and easy to use.
[0049] Furthermore, the difference filter residual module includes a pooling layer, two convolutional layers, and an upsampling unit;
[0050] The output end of the pooling layer is connected to the input ends of the two convolutional layers respectively. The output feature maps of the two convolutional layers are subtracted and input into the upsampling unit. The output edge information of the upsampling unit is multiplied with the original image to obtain enhanced edge information.
[0051] The difference filter residual module is conducive to enhancing edge information and is easy to use. BRIEF DESCRIPTION OF THE DRAWINGS
[0052] Figure 1 It is a flow chart of a target detection method based on multi-scale tokens of the present invention;
[0053] Figure 2 is a schematic diagram of the structure of a multi-scale internal token module of a target detection method based on multi-scale tokens of the present invention;
[0054] Figure 3 It is a structural schematic diagram of obtaining multi-scale output of the target detection method based on multi-scale tokens of the present invention;
[0055] Figure 4 It is a structural schematic diagram of the difference filter residual module of the target detection method based on multi-scale tokens of the present invention. DETAILED DESCRIPTION
[0056] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.
[0057] In the description of the present invention, it is necessary to understand that the terms "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.
[0058] In the description of the present invention, unless otherwise specified and limited, it should be noted that the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a mechanical connection or an electrical connection, or it can be the internal connection between two components. It can be a direct connection or an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to the specific circumstances.
[0059] The present invention discloses a target detection method based on multi-scale tokens, such as Figure 1 As shown, the following steps are included:
[0060] Get the original image and convolve it into two image blocks of size a and b; use the internal multi-scale token module to promote self-attention learning by merging information. In the first stage, H*W*3 is taken as the input image.
[0061] Construct an object detection network, which includes multiple multi-scale internal token modules (IMTs) of different scales and multiple corresponding difference filter residual modules (DFRs); IMT can enrich fine-grained features by focusing on multi-scale features within a single input token, effectively extracting more multi-scale features before the self-attention layer. DRF, as a convolutional residual network, helps to enhance the local spatial features of a single token and learn more edge information.
[0062] The image patch is input into the multi-scale internal token module and the difference filter residual module of the first stage;
[0063] The multi-scale output of the multi-scale internal token module in the previous stage and the output of the difference filter residual module in the previous stage are used as the input of the multi-scale internal token module in the next stage, and the output of the difference filter residual module in the previous stage is used as the input of the difference filter residual module in the next stage;
[0064] The output of the multi-scale internal token module in the final stage is fused with the output of the difference filter residual module in the final stage to obtain the target detection result.
[0065] In a preferred embodiment of the present invention, Figure 2 and Figure 3 As shown, Figure 3 middle:
[0066] DFR is a filtering module; Linear is a linear layer; Non-Linear is a nonlinear layer; DualFeature is a dual-branch feature extraction layer, which uses different convolution kernel sizes to convolve different feature maps; PatchEmbedding is used to convert the original 2D image into a series of 1D patch embeddings; Multi-sclae is a multi-scale operation; Concat is splicing, and different sizes need to be spliced after multi-scale; Relu is an activation function, which can achieve a nonlinear activation effect; Pool is a pooling layer, which realizes nonlinearity and can prevent overfitting to a certain extent.
[0067] The method for the multi-scale internal token module to obtain multi-scale output is as follows:
[0068] The image tokens are mapped to three matrices Q, K, and V. Convolution is used to downsample the matrices K and V to different scales, focusing attention at different levels, which is conducive to learning the underlying fine-grained features within a single token. IMT receives a query (Q), which is the spatial reduction of K and V to reduce computational / memory overhead. The details of IMT in stage I can be expressed as follows:
[0069]
[0070] K i,j =MS j (K i ), j=1,2,3...
[0071] V i,j =MS j (V i ), j=1,2,3...
[0072]
[0073] Among them, MS is a multi-scale function, i represents the sampling stage, j represents different scales; Q is the query matrix, K represents the matrix of the correlation between the queried information and other information, V represents the matrix of the queried information; W Qi , W Ki , W ViIndicates that a corresponding weight matrix is assigned to the feature map X; X is the feature map, and the attention weight of X corresponding to the information V is proportional to Q*K.tranpose, which means that the attention weight of X is determined by X itself, so it is called self-attention; W q ,W k ,W v It will be updated according to the task objectives to ensure the effectiveness of the self-attention mechanism; W is the weight matrix; Softmax is a normalization function, and the calculation results are distributed in the interval [0, 1]; Atttention is the symbolic mark of self-attention; T represents transposition, d k Represents the dimension of matrix K; MS j (K i ) indicates that the matrix K is transformed to the jth scale in the i-th stage; MS j (V i ) means that in the i-th stage, the matrix V is transformed to the j-scale size, Q i , K i , V i The query matrix, correlation matrix, and query information matrix calculated in the i-th sampling stage; the mean of the matrix product of Q and K should be 0, and the variance should be 1, so that a smoother softmax can be obtained. When the dimension is large, the dot product result will be large, resulting in a small softmax gradient. To mitigate this effect, the dot product is scaled.
[0074] Through multi-scale transformation, we can get matrices K of different sizes. i,j and V i,j , by upsampling the matrix K i,j and V i,j The matrix sizes are unified and then spliced;
[0075] For the matrix K i,j and V i,j Rearrange the elements in the row and column directions with a preset value (such as 2) as the interval, which is K j and V j Form new key-value pairs, expand the receptive domain, and achieve global perception without losing information;
[0076] A nonlinear transformation layer is applied to expand the search feature subspace. The nonlinear transformation layer includes a linear layer, a relu layer, a pooling layer, and a linear layer connected in sequence to fit richer image features. Combined with the linear transformation layer, more combinations of Q, K, and V are found to mine the fine-grained features of the mark. The linear transformation layer uses a fully connected layer to connect the transformation. The nonlinear transformation layer and the linear transformation layer are composed of two fully connected layers, a pooling layer, and an activation function:
[0077] Qnon =f(Activate(f(Q))+b)
[0078] K non =f(Activate(f(K))+b)
[0079] V non =f(Activate(f(V))+b)
[0080] Among them, f is a linear function, Activate is an activation function, and b is a bias;
[0081] Using the dual feature extraction method, the original image is convolved with different convolution kernels to obtain different feature maps to enrich the spatial context and fine-grained features of the input tag. For the input original image X, a non-overlapping sliding window of size 1*1 is used to obtain the embedded tag, focusing on the fine-grained features; an overlapping sliding window of size 3*3 and a preset stride (such as 2) is used to capture the spatial context.
[0082] In a preferred embodiment of the present invention, Figure 4 As shown in the figure, Avgpool is a pooling layer that reduces dimension and removes redundant information, while increasing the receptive field, retaining the feature information of the feature map and reducing the number of parameters; Cov is a convolution layer that convolves a feature map; subtracting two feature maps to obtain edge information; Upsampling is because it must be restored to its original size before it can be operated; multiplication operation multiplies the edge information obtained by the original image to enhance the edge information. The method for extracting high-frequency information at the input end by the difference filter residual module is as follows:
[0083] Assume the size of the image token is C*H*w, apply average pooling to obtain smooth features, and then use two convolution kernels to obtain two different smooth images G A , G B :
[0084] G A ,G B = Cov(Avgpool(G))
[0085] Among them, Cov represents the convolution function, the convolution kernel size is 3*3, the stride is 1, and the padding is 1; G is the original graph; Avgpool represents the pooling function, which performs average pooling on the original graph G, and the average pool kernel size is 2*2 and the stride is 2;
[0086] G A , G B Subtract the difference image, and then upsample it to restore the size of the feature map.
[0087] In a preferred embodiment of the present invention, the method for obtaining the target detection result is as follows:
[0088] Calculate and fuse based on the semantic importance of the edge information of the image:
[0089]
[0090]
[0091] Among them, f is a linear function that is similar to a 1*1 convolutional layer; T is the input feature map matrix, H and w are the height and width of the feature map, Gaus represents the target detection function, which is used to calculate edge information and the weight of each point, such as 0.2*T[12,68]; ζ is the normalization function, and x is the variable of the function, that is, the value in the matrix pixel, which is limited to between 0 and 1. Sum each element of the matrix and then find the average. Then return the obtained value to the interval of 0-1, and then multiply it by the original view feature map T to give the feature map a weight.
[0092] The present invention also provides a target detection system based on multi-scale tokens, including a data acquisition module and a processing module. The data acquisition module is used to acquire the original image. The output end of the data acquisition module is connected to the input end of the processing module. The processing module executes the method of the present invention to perform target detection.
[0093] In a preferred embodiment of the present invention, the processing module includes a plurality of multi-scale internal token modules of different scales and a plurality of differential filter residual modules corresponding thereto. The output end of the data acquisition module is respectively connected to the input end of the multi-scale internal token module of the first stage and the differential filter residual module of the first stage. The multi-scale output of the multi-scale internal token module of the previous stage and the output of the differential filter residual module of the previous stage are used as the input of the multi-scale internal token module of the next stage, and the output of the differential filter residual module of the previous stage is used as the input of the differential filter residual module of the next stage. The output of the multi-scale internal token module of the last stage and the output of the differential filter residual module of the last stage are fused to output the target detection result.
[0094] Preferably, Figure 3 As shown, the multi-scale internal token module includes a multi-scale operation unit, a splicing unit, a linear layer and a non-linear layer. The output end of the multi-scale operation unit is electrically connected to the input end of the splicing unit, and the output end of the splicing unit is electrically connected to the input ends of the linear layer and the non-linear layer respectively.
[0095] like Figure 4As shown, the difference filter residual module includes a pooling layer, two convolutional layers, and an upsampling unit. The output end of the pooling layer is connected to the input ends of the two convolutional layers respectively, and the output feature maps of the two convolutional layers are subtracted and input to the upsampling unit. The output edge information of the upsampling unit is multiplied with the original image to obtain enhanced edge information.
[0096] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.
[0097] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.
Claims
1. A target detection method based on multi-scale tokens, characterized in that: The steps include: Get the original image and convolve it into image blocks of two sizes, a and b; Constructing a target detection network, the target detection network includes a plurality of multi-scale internal token modules of different scales and a plurality of corresponding difference filter residual modules; The image patch is input into the multi-scale internal token module and the difference filter residual module of the first stage; The multi-scale output of the multi-scale internal token module in the previous stage and the output of the difference filter residual module in the previous stage are used as the input of the multi-scale internal token module in the next stage, and the output of the difference filter residual module in the previous stage is used as the input of the difference filter residual module in the next stage; The output of the multi-scale internal token module in the last stage is fused with the output of the difference filter residual module in the last stage to obtain the target detection result; The method of extracting high-frequency information from the input end of the difference filter residual module is as follows: Assume the size of the image token is C*H*w, apply average pooling to obtain smooth features, and then use two convolution kernels to obtain two different smooth images G A , G B : G A ,G B =Cov(Avgpool(G)) Among them, Cov represents the convolution function, the convolution kernel size is 3*3, the stride is 1, and the padding is 1; Avgpool represents the pooling function, the average pool kernel size is 2*2, and the stride is 2; G is the original graph; G A , G B Subtract to get the difference image, then upsample to restore the size of the feature map; The method to obtain the target detection result is as follows: Calculate and fuse based on the semantic importance of the edge information of the image: Among them, f is a linear function that is similar to a 1*1 convolutional layer; T is the input feature map matrix, H and w are the height and width of the feature map, Gaus represents the target detection function, ζ is the normalization function, and x is the variable of the function, that is, the matrix pixel value.
2. The target detection method based on multi-scale tokens as claimed in claim 1, characterized in that: The method for the multi-scale internal token module to obtain multi-scale output is as follows: The image tokens are mapped to three matrices Q, K, and V. The matrices K and V are downsampled to different scales using convolution: K i,j =MS j (K i ),j=1,2,3... V i,j =MS j (V i ),j=1,2,3... Among them, MS is a multi-scale function, i represents the sampling stage, j represents different scales; Q is the query matrix, K represents the matrix of the correlation between the queried information and other information, V represents the matrix of the queried information; X is the feature map, W is the weight matrix; W Qi , W Ki , W Vi Indicates that a corresponding weight matrix is assigned to the feature map X; Softmax is a normalization function, Attention is the symbolic mark of self-attention; T represents transposition, d k Represents the dimension of matrix K; MS j (K i ) indicates that in the i-th stage, the matrix K i Make j scale transformation, MS j (V i ) represents the matrix V in the i-th stage i Make a transformation of j scale, Q i , K i , V i The query matrix, correlation matrix, and query information matrix calculated in the i-th sampling stage; matrices K of different sizes are obtained through multi-scale transformation i,j and V i,j , by upsampling the matrix K i,j and V i,j The matrix sizes are unified and then spliced; For the matrix K i,j and V i,j Rearrange the elements in rows and columns with a preset interval, K j and V j Form new key-value pairs, expand the receptive domain, and achieve global perception without losing information; A nonlinear transformation layer is applied to expand the search feature subspace. The nonlinear transformation layer includes a linear layer, a relu layer, a pooling layer, and a linear layer connected in sequence to fit richer image features. Combined with the linear transformation layer, more combinations of Q, K, and V are found to mine the fine-grained features of the mark. The linear transformation layer uses a fully connected layer to connect the transformation. The nonlinear transformation layer consists of two fully connected linear transformation layers, a pooling layer, and an activation function: Q non =f(Activate(f(Q))+b) K non =f(Activate(f(K))+b) V non =f(Activate(f(V))+b) Among them, f is a linear function, Activate is an activation function, and b is a bias; Using the dual feature extraction method, the original image is convolved with different convolution kernels to obtain different feature maps to enrich the spatial context and fine-grained features of the input tag. For the input original image X, a non-overlapping sliding window of size 1*1 is used to obtain the embedded tag, focusing on the fine-grained features; an overlapping sliding window of size 3*3 and a preset stride is used to capture the spatial context.
3. A target detection system based on multi-scale tokens, characterized in that: It comprises a data acquisition module and a processing module, wherein the data acquisition module is used to acquire the original image, the output end of the data acquisition module is connected to the input end of the processing module, and the processing module executes the method according to claim 1 or 2 to perform target detection.
4. The multi-scale token-based target detection system according to claim 3, characterized in that: The processing module includes a plurality of multi-scale internal token modules of different scales and a plurality of differential filter residual modules corresponding thereto; The output end of the data acquisition module is respectively connected to the input end of the multi-scale internal token module of the first stage and the input end of the difference filter residual module of the first stage; The multi-scale output of the multi-scale internal token module in the previous stage and the output of the difference filter residual module in the previous stage are used as the input of the multi-scale internal token module in the next stage, and the output of the difference filter residual module in the previous stage is used as the input of the difference filter residual module in the next stage; The output of the multi-scale internal token module in the final stage is fused with the output of the difference filter residual module in the final stage to output the target detection result.
5. The multi-scale token-based target detection system according to claim 4, characterized in that: The multi-scale internal token module includes a multi-scale operation unit, a splicing unit, a linear layer and a non-linear layer, the output end of the multi-scale operation unit is connected to the input end of the splicing unit, and the output end of the splicing unit is connected to the input ends of the linear layer and the non-linear layer respectively.
6. The multi-scale token-based target detection system according to claim 4, characterized in that: The difference filter residual module includes a pooling layer, two convolutional layers, and an upsampling unit; The output end of the pooling layer is connected to the input ends of the two convolutional layers respectively. The output feature maps of the two convolutional layers are subtracted and input into the upsampling unit. The output edge information of the upsampling unit is multiplied with the original image to obtain enhanced edge information.