A learning-based edge detection method imitating the parallel and hierarchical mechanism of biological vision

Abstract

The present invention aims to provide a learning-based edge detection method that mimics the parallel and hierarchical mechanisms of biological vision, comprising the following steps: A. constructing an encoding network and a decoding network; wherein the encoding network includes a parallel processing network and a hierarchical processing network; B. inputting the original image into the parallel processing network, where it is processed by the X-cell sub-network and the Y-cell sub-network, respectively. The processing results are added and fused, and the resulting fused results are respectively input into the hierarchical processing network and the decoding network; C. The fused results are processed by the simple cell sub-network in the hierarchical processing network, resulting in simple cell processing results, which are respectively input into the complex cell sub-network and the decoding network; the complex cell sub-network processes the complex cell processing results, which are then input into the decoding network; and D. decoding by the decoding network to obtain the final detection result. This invention can achieve competitive performance while using very few parameters.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and in particular to a learning-type edge detection method that imitates the parallel and hierarchical mechanism of biological vision. Background Art

[0002] Edge detection, as a crucial foundation for advanced computer vision tasks, requires reserving as much computing resources as possible for these tasks while ensuring adequate performance. Convolutional neural network (CNN) encoder-decoder edge detection methods have developed rapidly, primarily using VGG16 or ResNet as encoders and achieving excellent performance through decoder design. These methods have demonstrated excellent performance in edge detection tasks. However, other edge detection methods rarely consider the implications of biological vision mechanisms for neural networks, leading to challenges in feature extraction and edge information extraction. These issues include weak feature extraction capabilities and insufficient extraction of edge information.

[0003] Current biological methods for simulating biological visual systems rely primarily on mathematical formulas to simulate certain characteristics of the biological visual system. However, this approach fails to effectively simulate most of the characteristics and information transmission mechanisms of the biological visual system, and therefore cannot effectively extract the edges of objects. Summary of the Invention

[0004] The present invention aims to provide a learning-based edge detection method that mimics the parallel and hierarchical mechanism of biological vision. This method can achieve competitive performance with minimal parameters and has higher parameter utilization efficiency, which can reserve more computing and storage resources for other visual processing tasks on a resource-constrained platform.

[0005] The technical solutions of the present invention are as follows:

[0006] The learning-based edge detection method that mimics the parallel and hierarchical mechanism of biological vision comprises the following steps:

[0007] A. Construct encoding network and decoding network;

[0008] Among them, the encoding network includes a parallel processing network and a hierarchical processing network;

[0009] The parallel processing network includes X cell sub-network and Y cell sub-network; the hierarchical processing network includes simple cell sub-network and complex cell sub-network;

[0010] B. The original image is input into the parallel processing network and processed by the X-cell sub-network and the Y-cell sub-network respectively. The processing results of the X-cell sub-network and the Y-cell sub-network are added and fused together. The obtained addition and fusion results are input into the hierarchical processing network and the decoding network respectively.

[0011] C. The additive fusion results are processed by the simple cell sub-network in the hierarchical processing network, and the simple cell processing results are input into the complex cell sub-network and the decoding network respectively; the complex cell sub-network processes the complex cell processing results and inputs them into the decoding network;

[0012] D. The decoding network decodes the addition fusion results, simple cell processing results and complex cell processing results to obtain the final detection results.

[0013] In step B, the processing process in the X cell subnetwork is as follows: the input result is processed by 7×7 depth-separable convolution and 21×21 depth-separable convolution respectively, the 7×7 depth-separable convolution result is subtracted from the 21×21 depth-separable convolution result, and then the output result is obtained by 1×1 convolution.

[0014] In the X cell sub-network, the central response mechanism is a 7×7 depth-separable convolutional layer Conv C The peripheral response mechanism is composed of a 21×21 depth separable convolution layer Conv s The effective receptive field is constructed into a ring, and an antagonistic mechanism is formed by subtraction.

[0015] The X-cell sub-network processing process is as follows:

[0016] The central response mechanism is represented by a 7×7 depth-separable convolution, and the calculation formula is as follows:

[0017]

[0018] Where i∈{1, 2, ..., n} is the channel number, It represents the characteristics of the i-th channel after the central mechanism, is the convolution kernel ConvC that simulates the central mechanism, ReLU(·) is the activation function, and Norm(·) is the normalization function;

[0019] The 21×21 depthwise separable convolution represents the peripheral response mechanism, and the calculation formula is as follows:

[0020]

[0021] Where i∈{1, 2, ..., n} is the channel number, represents the characteristics of the i-th channel after the peripheral mechanism, and is the convolution kernel (Convs) that simulates the peripheral mechanism, ReLU(·) is the activation function, and Norm(·) is the normalization function;

[0022] The calculation formula of the X cell sub-network is as follows:

[0023]

[0024] Where Ei represents the antagonistic response of the i-th channel, and ω is a learnable weight parameter that controls the inhibitory strength of the peripheral mechanism on the central mechanism;

[0025] After obtaining the X cell antagonistic response of each channel, the channels are integrated using 1×1 convolution to obtain the final X cell response.

[0026] In step B, the processing process in the Y cell sub-network is as follows:

[0027] The Y-cell model simulates the center-periphery mechanism in the same way as the X-cell model, except that the Y-cell model has a unique nonlinear subunit model.

[0028] The Y cell sub-network processing process is as follows:

[0029] The input result is divided into three paths. The first path undergoes 7×7 depth-separable convolution processing, the second path undergoes 21×21 depth-separable convolution processing, and the third path undergoes 3×3 depth-separable convolution processing and 21×21 depth-separable convolution processing in sequence; the first convolution result is added to the third convolution result, the second convolution result is subtracted, and then the output result is obtained by 1×1 convolution processing.

[0030] The central response mechanism is represented by a 7×7 depth-separable convolution, and the calculation formula is as follows:

[0031]

[0032] Where i∈{1, 2, ..., n} is the channel number, It represents the characteristics of the i-th channel after the central mechanism, is the convolution kernel ConvC that simulates the central mechanism, ReLU(·) is the activation function, and Norm(·) is the normalization function;

[0033] The 21×21 depthwise separable convolution represents the peripheral response mechanism, and the calculation formula is as follows:

[0034]

[0035] Where i∈{1, 2, ..., n} is the channel number, represents the characteristics of the i-th channel after the peripheral mechanism, and It is a convolution kernel that simulates the surrounding mechanism (Conv s ), ReLU(·) is the activation function, and Norm(·) is the normalization function;

[0036] Nonlinear subunit response

[0037] The influence range of the nonlinear subunit is divided into multiple concentric circle areas according to the spacing w, that is, multiple 21×21 depth-separable convolutions of different sizes. col By adjusting the radius and controlling the size of the ring, each area is responsible for a nonlinear subunit, which is composed of a 3×3 depth-separable convolution Conv sub constitute;

[0038] It represents the characteristics of the i-th channel after the nonlinear subunit response, and its calculation formula is as follows:

[0039]

[0040] Where k represents the number of nonlinear subunits, is the learnable convolution kernel of the jth nonlinear subunit, It is responsible for collecting the response of the j-th nonlinear subunit in the responsible area to the central position. Its shape is circular with a width of w, and it forms a series connection with the convolution operation of the nonlinear subunit.

[0041] The calculation formula of the Y cell sub-network is as follows:

[0042]

[0043] Where Ei represents the antagonistic response of the i-th channel, and ω is a learnable weight parameter that controls the inhibitory strength of the peripheral mechanism on the central mechanism;

[0044] After obtaining the Y cell responses of each channel, 1×1 convolution is used to integrate each channel to obtain the final Y cell response.

[0045] In step C, the simple cell sub-network is simulated by using four-directional learnable direction-selective convolution kernels, namely horizontal, vertical, left-leaning, and right-leaning. Based on this, for an input feature F containing n channels, the simple cell response of its i-th channel is The formula is:

[0046]

[0047] MaxPool(·) is a 2×2 pooling operation used to increase the receptive field of simple cells. is the learnable convolution kernel of a simple cell in the jth direction;

[0048] After obtaining the simple cell response of each channel based on the above formula, each channel is integrated using 1×1 convolution to obtain the final simple cell response.

[0049] In step C, the complex cell sub-network is simulated by using four-directional learnable direction-selective convolution kernels, namely horizontal, vertical, left-leaning, and right-leaning. Based on this, for an input feature F containing n channels, the complex cell response of its i-th channel is The formula is:

[0050]

[0051] MaxPool(·) is a 2×2 pooling operation used to increase the receptive field of simple cells. is the learnable convolution kernel of a simple cell in the jth direction;

[0052] After calculating the complex cell response of each channel based on the above formula, the channels are integrated using 1×1 convolution to obtain the final complex cell response.

[0053] The processing process in the decoding network is as follows:

[0054] The addition fusion result and the simple cell processing result are respectively input into the first FIM module for processing, and the processing result of the first FIM module and the complex cell processing result are respectively input into the second FIM module for processing. The processing result of the second FIM module is subjected to 1×1 convolution to obtain the final detection result.

[0055] The first FIM module and the second FIM module have the same structure, and the processing process is as follows:

[0056] The low-resolution input result is processed by 3×3 convolution, Norm function, and ReLU function in sequence, and then the resolution is consistent with the high-resolution input result. Then, it is processed by the MSA module to obtain the first MSA module processing result;

[0057] The high-resolution input result is processed by 3×3 convolution, Norm function, ReLU function, and MSA module in sequence to obtain the processing result of the second MSA module;

[0058] The output result is obtained by adding and fusing the processing result of the first MSA module and the processing result of the second MSA module.

[0059] The processing process in the MSA module is as follows:

[0060] After the input result undergoes 1×1 convolution, it is divided into four paths. The first path undergoes 9×9 convolution, the second path undergoes 7×7 convolution, the third path undergoes 5×5 convolution, and the fourth path undergoes 3×3 convolution. After the four-path processing results are added and fused, they are processed in sequence by 3×3 convolution, 1×1 convolution, and Sigmoid function.

[0061] The output result is obtained by multiplying the Sigmoid function processing result and the input result.

[0062] Guided by the visual information processing mechanism of parallel and hierarchical processing in the visual cortex, this paper combines CNNs to design a lightweight edge detection network, PHNet. This network simulates the visual information transmission pathway from the retina to the LGN and then to the V1 region, designing an encoding network with both parallel and hierarchical processing. The receptive fields of the cells involved in this pathway are carefully modeled. Experiments demonstrate that this method achieves excellent edge detection performance at a low computational cost.

[0063] This paper, based on the principle of sacrificing a small amount of network performance in exchange for lower computational cost, proposes a lightweight edge detection deep learning model that simulates the parallel hierarchical visual information processing mechanism of visual cortical neurons. Large-scale learnable convolution kernels are used to more closely simulate the shape and response mechanism of the cell's receptive field. Circular and annular learnable convolution kernels are used to simulate the center-periphery mechanism of the receptive fields of X and Y cells, and the antagonistic nature of the convolution responses is reflected through subtraction. Small convolution kernels and annular learnable convolution kernels are used to simulate the properties of the nonlinear subunits of Y cells, and a cascaded combination is used to realize the contribution of nonlinear subunits to the center mechanism. Finally, narrow and long learnable convolution kernels oriented in different directions are used to simulate the directional selectivity of the receptive field of simple cells.

[0064] Compared with some traditional edge detection networks, the method of the present invention can demonstrate competitive performance while using very few parameters, and has higher parameter utilization efficiency. It can reserve more computing resources and storage resources for other visual processing tasks on a resource-compact platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 This is a schematic diagram of the overall network structure of Example 1 of the present invention;

[0066] Figure 2 Schematic diagram of the X-cell subnetwork of Example 1, wherein: (a) simulation of the X-cell receptive field; (b) ranges of central and peripheral mechanisms; (c) X-cell subnetwork;

[0067] Figure 3 Schematic diagram of the Y cell subnetwork of Example 1, including: (a) Y cell receptive field simulation; (b) nonlinear subunit responsibility area division method; (c) Y cell subnetwork;

[0068] Figure 4 Schematic diagram of four convolution kernels simulating direction-selective receptive fields in Example 1;

[0069] Figure 5This is a schematic diagram of the structure of the hierarchical processing network of Example 1;

[0070] Figure 6 Schematic diagram of the decoding network structure;

[0071] Figure 7 This is a comparison chart of the contour detection effects of the solution of the present invention in Example 2 and the solution of Reference 1. DETAILED DESCRIPTION

[0072] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0073] Example 1

[0074] The learning-based edge detection method that mimics the parallel and hierarchical mechanism of biological vision comprises the following steps:

[0075] A. Construct Figure 1-3 , the encoding network and decoding network shown in 5-6;

[0076] Among them, the encoding network includes a parallel processing network and a hierarchical processing network;

[0077] The parallel processing network includes X cell sub-network and Y cell sub-network; the hierarchical processing network includes simple cell sub-network and complex cell sub-network;

[0078] B. The original image is input into the parallel processing network and processed by the X-cell sub-network and the Y-cell sub-network respectively. The processing results of the X-cell sub-network and the Y-cell sub-network are added and fused together. The obtained addition and fusion results are input into the hierarchical processing network and the decoding network respectively.

[0079] The processing process in the X-cell subnetwork is as follows: the input results are processed by 7×7 depthwise separable convolution and 21×21 depthwise separable convolution respectively, the 7×7 depthwise separable convolution result is subtracted from the 21×21 depthwise separable convolution result, and then the output result is processed by 1×1 convolution. The details are as follows:

[0080] The central response mechanism is represented by a 7×7 depth-separable convolution, and the calculation formula is as follows:

[0081]

[0082] Where i∈{1, 2, ..., n} is the channel number, It represents the characteristics of the i-th channel after the central mechanism, is the convolution kernel ConvC that simulates the central mechanism, ReLU(·) is the activation function, and Norm(·) is the normalization function;

[0083] The 21×21 depthwise separable convolution represents the peripheral response mechanism, and the calculation formula is as follows:

[0084]

[0085] Where i∈{1, 2, ..., n} is the channel number, represents the characteristics of the i-th channel after the peripheral mechanism, and is the convolution kernel (Convs) that simulates the peripheral mechanism, ReLU(·) is the activation function, and Norm(·) is the normalization function;

[0086] The calculation formula of the X cell sub-network is as follows:

[0087]

[0088] Where Ei represents the antagonistic response of the i-th channel, and ω is a learnable weight parameter that controls the inhibitory strength of the peripheral mechanism on the central mechanism;

[0089] After obtaining the X cell antagonistic response of each channel, the channels are integrated using 1×1 convolution to obtain the final X cell response.

[0090] The processing process in the Y cell sub-network is as follows:

[0091] The input result is divided into three paths. The first path is processed by 7×7 depth-separable convolution, the second path is processed by 21×21 depth-separable convolution, and the third path is processed by 3×3 depth-separable convolution and 21×21 depth-separable convolution in sequence. The first convolution result is added to the third convolution result, and the second convolution result is subtracted. Then, the output result is processed by 1×1 convolution. The details are as follows:

[0092] The central response mechanism is represented by a 7×7 depth-separable convolution, and the calculation formula is as follows:

[0093]

[0094] Where i∈{1, 2, ..., n} is the channel number, It represents the characteristics of the i-th channel after the central mechanism, is the convolution kernel ConvC that simulates the central mechanism, ReLU(·) is the activation function, and Norm(·) is the normalization function;

[0095] The 21×21 depthwise separable convolution represents the peripheral response mechanism, and the calculation formula is as follows:

[0096]

[0097] Where i∈{1, 2, ..., n} is the channel number, represents the characteristics of the i-th channel after the peripheral mechanism, and is the convolution kernel (Convs) that simulates the peripheral mechanism, ReLU(·) is the activation function, and Norm(·) is the normalization function;

[0098] Nonlinear subunit response

[0099] The influence range of the nonlinear subunit is divided into multiple concentric circle areas according to the spacing w, that is, multiple 21×21 depth-wise separable convolutions Convcol of different sizes. The size of the circle is controlled by adjusting the radius. Each area is responsible for a nonlinear subunit, which is composed of a 3×3 depth-wise separable convolution Convsub.

[0100] It represents the characteristics of the i-th channel after the nonlinear subunit response, and its calculation formula is as follows:

[0101]

[0102] Among them, boat represents the number of nonlinear subunits, is the learnable convolution kernel of the jth nonlinear subunit, It is responsible for collecting the response of the j-th nonlinear subunit in the responsible area to the central position. Its shape is circular with a width of w, and it forms a series connection with the convolution operation of the nonlinear subunit.

[0103] The calculation formula of the Y cell sub-network is as follows:

[0104]

[0105] Where Ei represents the antagonistic response of the i-th channel, and ω is a learnable weight parameter that controls the inhibitory strength of the peripheral mechanism on the central mechanism;

[0106] After obtaining the Y cell responses of each channel, 1×1 convolution is used to integrate each channel to obtain the final Y cell response.

[0107] C. The additive fusion results are processed by the simple cell sub-network in the hierarchical processing network, and the simple cell processing results are input into the complex cell sub-network and the decoding network respectively; the complex cell sub-network processes the complex cell processing results and inputs them into the decoding network;

[0108] The simple cell sub-network uses four strip-shaped learnable direction-selective convolution kernels to simulate it, such as Figure 4 As shown, the four directions are horizontal, vertical, left tilt, and right tilt. According to this, for an input feature F containing n channels, the simple cell response of its i-th channel is The formula is:

[0109]

[0110] MaxPool(·) is a 2×2 pooling operation used to increase the receptive field of simple cells. is the learnable convolution kernel of a simple cell in the jth direction;

[0111] After obtaining the simple cell response of each channel based on the above formula, each channel is integrated using 1×1 convolution to obtain the final simple cell response.

[0112] The complex cell sub-network is simulated using four-directional strip-shaped learnable direction-selective convolution kernels, such as Figure 4 As shown, the four directions are horizontal, vertical, left tilt, and right tilt. According to this, for an input feature F containing n channels, the complex cell response of its i-th channel is The formula is:

[0113]

[0114] MaxPool(·) is a 2×2 pooling operation used to increase the receptive field of simple cells. is the learnable convolution kernel of a simple cell in the jth direction;

[0115] After calculating the complex cell response of each channel based on the above formula, the channels are integrated using 1×1 convolution to obtain the final complex cell response.

[0116] D. The decoding network decodes the addition fusion results, simple cell processing results and complex cell processing results to obtain the final detection results.

[0117] The processing process in the decoding network is as follows:

[0118] The addition fusion result and the simple cell processing result are respectively input into the first FIM module for processing, and the processing result of the first FIM module and the complex cell processing result are respectively input into the second FIM module for processing. The processing result of the second FIM module is subjected to 1×1 convolution to obtain the final detection result.

[0119] The first FIM module and the second FIM module have the same structure, and the processing process is as follows:

[0120] The low-resolution input result is processed by 3×3 convolution, Norm function, and ReLU function in sequence, and then the resolution is consistent with the high-resolution input result. Then, it is processed by the MSA module to obtain the first MSA module processing result;

[0121] The high-resolution input result is processed by 3×3 convolution, Norm function, ReLU function, and MSA module in sequence to obtain the processing result of the second MSA module;

[0122] The output result is obtained by adding and fusing the processing result of the first MSA module and the processing result of the second MSA module.

[0123] The processing process in the MSA module is as follows:

[0124] After the input result undergoes 1×1 convolution, it is divided into four paths. The first path undergoes 9×9 convolution, the second path undergoes 7×7 convolution, the third path undergoes 5×5 convolution, and the fourth path undergoes 3×3 convolution. After the four-path processing results are added and fused, they are processed in sequence by 3×3 convolution, 1×1 convolution, and Sigmoid function.

[0125] The output result is obtained by multiplying the Sigmoid function processing result and the input result.

[0126] Example 2

[0127] For the quantitative performance evaluation of the final edge map, we use the same performance measurement criteria as in Reference 1. The specific evaluation is shown in the following formula.

[0128]

[0129] Where P represents precision and R represents recall. A larger value of F indicates better performance.

[0130] Where P represents precision and R represents recall. A larger value of F indicates better performance.

[0131] Document 1: Xie S, Tu Z. Holistically-Nested Edge Detection[J]. International Journal of Computer Vision, 2017, 125(1-3):3.

[0132] The parameters used in Reference 1 are the same as those in the original text, and are guaranteed to be the optimal parameters for the model.

[0133] Table 1 summarizes the experimental data of Reference 1 and Example 1 on the Berkeley segmentation dataset (BSDS500). From the experimental results, the method of the present invention shows competitive performance with very few parameters and has higher parameter utilization efficiency, which reserves more computing resources for other advanced tasks. Figure 7 .

[0134] Table 1 Performance comparison of the edge detection method provided by the embodiment and the edge detection method of reference 1

[0135]

Claims

1. A learning-based edge detection method that mimics the parallel and hierarchical mechanism of biological vision, characterized by The following steps are involved: A. Construct encoding network and decoding network; Among them, the encoding network includes a parallel processing network and a hierarchical processing network; The parallel processing network includes X cell sub-network and Y cell sub-network; the hierarchical processing network includes simple cell sub-network and complex cell sub-network; B. The original image is input into the parallel processing network and processed by the X-cell sub-network and the Y-cell sub-network respectively. The processing results of the X-cell sub-network and the Y-cell sub-network are added and fused together. The obtained addition and fusion results are input into the hierarchical processing network and the decoding network respectively. C. The additive fusion results are processed by the simple cell sub-network in the hierarchical processing network, and the simple cell processing results are input into the complex cell sub-network and the decoding network respectively; the complex cell sub-network processes the complex cell processing results and inputs them into the decoding network; D. The decoding network decodes the addition fusion results, simple cell processing results, and complex cell processing results to obtain the final detection results; In step B, the processing process in the X-cell subnetwork is as follows: the input result is processed by 7×7 depth-separable convolution and 21×21 depth-separable convolution respectively, the 7×7 depth-separable convolution result is subtracted from the 21×21 depth-separable convolution result, and then the output result is obtained by 1×1 convolution; The processing process in the decoding network is as follows: The addition fusion result and the simple cell processing result are respectively input into the first FIM module for processing. The processing result of the first FIM module and the complex cell processing result are respectively input into the second FIM module for processing. The processing result of the second FIM module is subjected to 1×1 convolution to obtain the final detection result. The first FIM module and the second FIM module have the same structure, and the processing process is as follows: The low-resolution input result is processed by 3×3 convolution, Norm function, and ReLU function in sequence, and then the resolution is consistent with the high-resolution input result. It is then processed by the MSA module to obtain the first MSA module processing result. The high-resolution input result is processed by 3×3 convolution, Norm function, ReLU function, and MSA module in sequence to obtain the processing result of the second MSA module; The output result is obtained by adding and fusing the processing result of the first MSA module and the processing result of the second MSA module; The processing process in the MSA module is as follows: After the input result undergoes 1×1 convolution, it is divided into four paths. The first path undergoes 9×9 convolution, the second path undergoes 7×7 convolution, the third path undergoes 5×5 convolution, and the fourth path undergoes 3×3 convolution. After the four-path processing results are added and fused, they are processed in sequence by 3×3 convolution, 1×1 convolution, and Sigmoid function. The output result is obtained by multiplying the Sigmoid function processing result and the input result.

2. The learning-based edge detection method based on the parallel and hierarchical mechanism of biological vision as claimed in claim 1, characterized in that: The X-cell sub-network processing process is as follows: The central response mechanism is represented by a 7×7 depth-separable convolution, and the calculation formula is as follows: Where i∈{1, 2, ..., n} is the channel number, It represents the characteristics of the i-th channel after the central mechanism, It is the convolution kernel Conv that simulates the central mechanism C , ReLU(·) is the activation function, Norm(·) is the normalization function; The 21×21 depthwise separable convolution represents the peripheral response mechanism, and the calculation formula is as follows: Where i∈{1, 2, ..., n} is the channel number, represents the characteristics of the i-th channel after the peripheral mechanism, and It is a convolution kernel that simulates the surrounding mechanism (Conv s ), ReLU(·) is the activation function, and Norm(·) is the normalization function; The calculation formula of the X cell sub-network is as follows: Among them, E i It is expressed as the antagonistic response of the i-th channel, and ω is a learnable weight parameter that controls the inhibitory strength of the peripheral mechanism on the central mechanism; After obtaining the X cell antagonistic response of each channel, the channels are integrated using 1×1 convolution to obtain the final X cell response.

3. The learning-based edge detection method based on the parallel and hierarchical mechanism of biological vision as claimed in claim 1, characterized in that: In step B, the processing process in the Y cell sub-network is as follows: The input result is divided into three paths. The first path undergoes 7×7 depth-separable convolution processing, the second path undergoes 21×21 depth-separable convolution processing, and the third path undergoes 3×3 depth-separable convolution processing and 21×21 depth-separable convolution processing in sequence; the first convolution result is added to the third convolution result, the second convolution result is subtracted, and then the output result is obtained by 1×1 convolution processing.

4. The learning-based edge detection method based on the parallel and hierarchical mechanism of biological vision as claimed in claim 3, characterized in that: The Y cell sub-network processing process is as follows: The central response mechanism is represented by a 7×7 depth-separable convolution, and the calculation formula is as follows: Where i∈{1, 2, ..., n} is the channel number, It represents the characteristics of the i-th channel after the central mechanism, It is the convolution kernel Conv that simulates the central mechanism C , ReLU(·) is the activation function, Norm(·) is the normalization function; The 21×21 depthwise separable convolution represents the peripheral response mechanism, and the calculation formula is as follows: Where i∈{1, 2, ..., n} is the channel number, represents the characteristics of the i-th channel after the peripheral mechanism, and It is the convolution kernel Conv that simulates the peripheral mechanism s , ReLU(·) is the activation function, Norm(·) is the normalization function; Nonlinear subunit response The influence range of the nonlinear subunit is divided into multiple concentric circle areas according to the spacing w, that is, multiple 21×21 depth-separable convolutions of different sizes. col By adjusting the radius and controlling the size of the ring, each area is responsible for a nonlinear subunit, which is composed of a 3×3 depth-separable convolution Conv sub constitute; It represents the characteristics of the i-th channel after the nonlinear subunit response, and its calculation formula is as follows: Where k represents the number of nonlinear subunits, is the learnable convolution kernel of the jth nonlinear subunit, It is responsible for collecting the response of the j-th nonlinear subunit in the responsible area to the central position. Its shape is circular with a width of w, and it forms a series connection with the convolution operation of the nonlinear subunit. The calculation formula of the Y cell sub-network is as follows: Among them, E i It is expressed as the antagonistic response of the i-th channel, and ω is a learnable weight parameter that controls the inhibitory strength of the peripheral mechanism on the central mechanism; After obtaining the Y cell responses of each channel, 1×1 convolution is used to integrate each channel to obtain the final Y cell response.

5. The learning-based edge detection method based on the parallel and hierarchical mechanism of biological vision as claimed in claim 1, characterized in that: In step C, the simple cell sub-network is simulated by using four-directional learnable direction-selective convolution kernels, namely horizontal, vertical, left-leaning, and right-leaning. Based on this, for an input feature F containing n channels, the simple cell response of its i-th channel is The formula is: MaxPool(·) is a 2×2 pooling operation used to increase the receptive field of simple cells. is the learnable convolution kernel of a simple cell in the jth direction; After obtaining the simple cell response of each channel based on the above formula, each channel is integrated using 1×1 convolution to obtain the final simple cell response.

6. The learning-based edge detection method based on the parallel and hierarchical mechanism of biological vision as claimed in claim 1, characterized in that: In step C, the complex cell sub-network is simulated by using four-directional learnable direction-selective convolution kernels, namely horizontal, vertical, left-leaning, and right-leaning. Based on this, for an input feature F containing n channels, the complex cell response of its i-th channel is The formula is: MaxPool(·) is a 2×2 pooling operation used to increase the receptive field of simple cells. is the learnable convolution kernel of a simple cell in the jth direction; After calculating the complex cell response of each channel based on the above formula, the channels are integrated using 1×1 convolution to obtain the final complex cell response.

Images