An edge detection method of a pseudo-parallel channel XYW cell

Abstract

The application aims to provide an edge detection method of a pseudo-parallel channel XYW cell, comprising the following steps: A, constructing a neural network, wherein the neural network comprises an encoding network and a decoding network; wherein the encoding network comprises sequentially connected Init, Path1, Path2, Path3, Path4 modules; B, inputting an original image into the encoding network, adjusting the channel number to C through the Init module first, inputting the original image into the Path1 module for processing and adjusting the channel number to 2C, inputting the original image into the Path2 module for processing and adjusting the channel number to 4C, inputting the original image into the Path3 module for processing and then inputting the original image into the Path4 module, and inputting the processing result of the Path4 module into the decoding network; C, decoding the processing result of the Path4 module in the decoding network to obtain a final detection result. The application can obtain excellent performance with low parameter quantity without the need of migration learning.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and in particular to an edge detection method for simulating parallel pathway XYW cells. Background Art

[0002] Traditional edge detection focuses on extracting edge information from discrete image features, such as changes in brightness, color, or texture. However, these operators cannot effectively distinguish between edge and texture information. To achieve better edge detection performance, some researchers have designed edge detection models using single functions to mimic biological vision mechanisms. Compared to traditional detection operators, these models exhibit better texture suppression capabilities, resulting in superior edge detection performance. However, relying on a single function is insufficient to fully represent the complex biological vision mechanisms. With the advancement of deep learning, CNNs have demonstrated powerful performance in various visual tasks, and encoder-decoder CNN models have also achieved significant breakthroughs in edge detection. However, these edge detection networks use the VGG16 encoder network, which was originally designed for classification tasks. These models have large parameters and a complex structure, requiring transfer learning during training to achieve better edge detection performance.

[0003] While CNN-based encoder-decoder edge detection methods have achieved remarkable results in recent years, most have focused on improving edge detection accuracy by using the VGG16 encoding network and designing the decoding network. However, using the VGG16 classification network as the encoding network presents challenges such as large model size, numerous parameters, and complex architecture. Furthermore, transfer learning is required during training to achieve better performance. Furthermore, the role of effective edge detection mechanisms in biological vision is neglected. Summary of the Invention

[0004] The present invention aims to provide an edge detection method that simulates parallel pathway XYW cells. This method is inspired by the effective edge detection mechanism and information transmission mode of biological vision, and can achieve excellent performance with a low number of parameters without the need for transfer learning.

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

[0006] The edge detection method of the parallel path XYW cell simulation comprises the following steps:

[0007] A. Constructing a neural network, wherein the neural network includes an encoding network and a decoding network;

[0008] The encoding network includes an Init module, a Path1 module, a Path2 module, a Path3 module, and a Path4 module connected in sequence. The number of channels of the Path1 module is C, the number of channels of the Path2 module is 2C, the number of channels of the Path3 module is 4C, and the number of channels of the Path4 module is 4C.

[0009] B. The original image is input into the encoding network. After the number of channels is adjusted to C by the Init module, it is input into the Path1 module. After being processed by the Path1 module and the number of channels is adjusted to 2C, it is input into the Path2 module. After being processed by the Path2 module and the number of channels is adjusted to 4C, it is input into the Path3 module. After being processed by the Path3 module, it is input into the Path4 module. The processing result of the Path4 module is input into the decoding network.

[0010] C. The decoding network decodes the processing results of the Path4 module to obtain the final detection results.

[0011] The basic structures of the Path1 module, Path2 module, Path3 module, and Path4 module are the same.

[0012] The processing procedures in the Path1 module, Path2 module, and Path3 module are the same, and the processing procedures are as follows:

[0013] The input results are divided into three paths. The first path is processed by three X-cell modules in sequence. When the first X-cell module is processed, the number of channels is doubled. The second path is processed by three Y-cell modules in sequence. When the first Y-cell module is processed, the number of channels is doubled. The third path is processed by three W-cell modules in sequence. When the first W-cell module is processed, the number of channels is doubled. The output result is obtained by adding and fusing the results of the three processes.

[0014] The processing process in the Path4 module is as follows: the input result is divided into three paths, the first path is processed in sequence by three X cell modules, the second path is processed in sequence by three Y cell modules, and the third path is processed in sequence by three W cell modules; the results of the three paths are added and fused to obtain the output result.

[0015] The X-cell module includes a 1×1 convolution, a 3×3 depth separation convolution, and a subtraction fusion module. The input result is divided into two paths, one path undergoes 1×1 convolution processing, and the other path undergoes 3×3 depth separation convolution processing; the two processing results are subtracted and fused to obtain the output result;

[0016] In the first X-cell module of the Path1 module, Path2 module, and Path3 module, the number of channels will be doubled during 1×1 convolution and 3×3 depthwise separation convolution processing.

[0017] The Y cell module includes a 1×1 convolution, a 5×5 depth-separable dilated convolution, and a subtractive fusion module. The input result is divided into two paths, one path undergoes 1×1 convolution processing, and the other path undergoes 5×5 depth-separable dilated convolution processing; the two processing results are subtracted and fused to obtain the output result;

[0018] In the first Y cell module of Path1 module, Path2 module, and Path3 module, the number of channels will be doubled during 1×1 convolution and 5×5 depthwise separable dilated convolution processing.

[0019] The W cell module includes 1×3 depth separation convolution and 3×1 depth separation convolution. The input is processed by 1×3 depth separation convolution and 3×1 depth separation convolution in sequence to obtain the output result;

[0020] In the first W cell module of the Path1 module, Path2 module, and Path3 module, the number of channels will be doubled during 1×3 depthwise separation convolution processing.

[0021] The Init module is a 7*7 dilated convolution with a dilation rate of 2.

[0022] The decoding network includes three ITM modules, 1×1 convolution, and Sigmoid function. The processing results of the Path4 module and the Path3 module are respectively input into the first ITM module, the processing results of the first ITM module and the Path2 module are respectively input into the second ITM module, the processing results of the second ITM module and the Path1 module are respectively input into the third ITM module, and the processing results of the third ITM module are sequentially processed by 1×1 convolution and Sigmoid function to obtain the final detection result.

[0023] The processing process in the ITM module is as follows: of the two input results, the input structure with high resolution is processed by the ELC module, normalization function, and ReLU function in sequence to obtain the first processing result; the input result with low resolution is processed by the ELC module, normalization function, and ReLU function in sequence, and then upsampled to the same resolution as the high-resolution input result to obtain the second processing result; the two processing results are added to obtain the output result.

[0024] The ELC module processing process is as follows: calling the convolution module of the pytorch framework, then constructing the second-order difference between the interval value around the convolution kernel and the center value, setting the intermediate convolution kernel value to zero, obtaining a new convolution module, and using the new convolution module to process the input result to obtain the output result.

[0025] The present invention simulates the characteristics of the receptive fields of X, Y, and W cells in the parallel pathway, designs three modules, X, Y, and W, and combines the three modules according to the parallel pathway to form a new encoding network, thereby achieving better edge feature extraction.

[0026] The present invention simulates the feature integration characteristics of the IT cortex and designs a new feature integration module in the decoding network to achieve better integration of edge features.

[0027] The method proposed in the present invention can effectively extract edge feature information, thereby obtaining a more accurate edge map. In addition, the network structure design and module design of the present invention are supported by physiological visual evidence, and achieve excellent performance with a low parameter count of 0.79M. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 Schematic diagram of the structure of the neural network of Example 1 of the present invention;

[0029] Figure 2 This is a schematic diagram of the structures of the Path1 module, Path2 module, Path3 module, and Path4 module of Example 1;

[0030] Figure 3 Schematic diagram of the X cell module of Example 1;

[0031] Figure 4 Schematic diagram of the Y cell module of Example 1;

[0032] Figure 5 This is a schematic diagram of the W cell module of Example 1;

[0033] Figure 6 It is a structural diagram of the ITM module;

[0034] Figure 7 It is a structural diagram of the ELC module;

[0035] Figure 8 This is a comparison chart of the contour detection effects of the solution of the present invention in Example 1 and the solution of Reference 1. DETAILED DESCRIPTION

[0036] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0037] Example 1

[0038] The edge detection method of the parallel path XYW cell simulation comprises the following steps:

[0039] A. Construct Figure 1 The neural network shown includes an encoding network and a decoding network;

[0040] The encoding network includes the Init module, Path1 module, Path2 module, Path3 module, and Path4 module connected in sequence; the number of channels of the Path1 module is set to C, the number of channels of the Path2 module is 2C, the number of channels of the Path3 module is 4C, and the number of channels of the Path4 module is 4C; the structures of the Path1 module, Path2 module, Path3 module, and Path4 module are shown in Figure 2 ;

[0041] B. The original image is input into the encoding network. After the number of channels is adjusted to C by the Init module, it is input into the Path1 module. After being processed by the Path1 module and the number of channels is adjusted to 2C, it is input into the Path2 module. After being processed by the Path2 module and the number of channels is adjusted to 4C, it is input into the Path3 module. After being processed by the Path3 module, it is input into the Path4 module. The processing result of the Path4 module is input into the decoding network.

[0042] The Init module is a 7*7 dilated convolution with a dilation rate of 2.

[0043] The processing procedures in the Path1 module, Path2 module, and Path3 module are the same, and the processing procedures are as follows:

[0044] The input results are divided into three paths. The first path is processed by three X-cell modules in sequence. When the first X-cell module is processed, the number of channels is doubled. The second path is processed by three Y-cell modules in sequence. When the first Y-cell module is processed, the number of channels is doubled. The third path is processed by three W-cell modules in sequence. When the first W-cell module is processed, the number of channels is doubled. The output result is obtained by adding and fusing the results of the three processes.

[0045] The processing process in the Path4 module is as follows: the input result is divided into three paths, the first path is processed in sequence by three X cell modules, the second path is processed in sequence by three Y cell modules, and the third path is processed in sequence by three W cell modules; the results of the three paths are added and fused to obtain the output result.

[0046] like Figure 3 As shown, the X cell module includes 1×1 convolution, 3×3 depth separation convolution, and subtraction fusion modules. The input result is divided into two paths, one path undergoes 1×1 convolution processing, and the other path undergoes 3×3 depth separation convolution processing; the two processing results are subtracted and fused to obtain the output result;

[0047] In the first X-cell module of the Path1 module, Path2 module, and Path3 module, the number of channels will be doubled during 1×1 convolution and 3×3 depthwise separation convolution processing.

[0048] like Figure 4 As shown, the Y cell module includes a 1×1 convolution, a 5×5 depth-separable dilated convolution, and a subtractive fusion module. The input result is divided into two paths, one path is processed by 1×1 convolution, and the other path is processed by 5×5 depth-separable dilated convolution. The two processing results are subtracted and fused to obtain the output result;

[0049] In the first Y cell module of Path1 module, Path2 module, and Path3 module, the number of channels will be doubled during 1×1 convolution and 5×5 depthwise separable dilated convolution processing.

[0050] like Figure 5 As shown, the W cell module includes 1×3 depth separation convolution and 3×1 depth separation convolution. The input is processed by 1×3 depth separation convolution and 3×1 depth separation convolution in sequence to obtain the output result;

[0051] In the first W cell module of the Path1 module, Path2 module, and Path3 module, the number of channels will be doubled during 1×3 depthwise separation convolution processing.

[0052] C. The decoding network decodes the results processed by the Path4 module. The specific decoding process is as follows:

[0053] The decoding network includes three ITM modules, 1×1 convolution, and Sigmoid function. The processing results of the Path4 module and the Path3 module are respectively input into the first ITM module, the processing results of the first ITM module and the Path2 module are respectively input into the second ITM module, the processing results of the second ITM module and the Path1 module are respectively input into the third ITM module, and the processing results of the third ITM module are sequentially processed by 1×1 convolution and Sigmoid function to obtain the final detection result.

[0054] like Figure 6 As shown, the processing process in the ITM module is as follows: of the two input results, the input structure with high resolution is processed by the ELC module, normalization function, and ReLU function in sequence to obtain the first processing result; the input result with low resolution is processed by the ELC module, normalization function, and ReLU function in sequence, and then upsampled to the same resolution as the high-resolution input result to obtain the second processing result; the two processing results are added together to obtain the output result.

[0055] like Figure 7As shown in FIG, the ELC module processing process is as follows: calling the convolution module of the pytorch framework, then constructing the second-order difference between the interval value around the convolution kernel and the center value, setting the intermediate convolution kernel value to zero, obtaining a new convolution module, and using the new convolution module to process the input result to obtain the output result.

[0056] Example 2

[0057] 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.

[0058]

[0059] Where P represents precision and R represents recall. A larger value of f indicates better performance.

[0060] Where P represents precision and R represents recall. A larger value of F indicates better performance.

[0061] In addition, we also consider the number of parameters as one of our performance indicators.

[0062] Document 1: Xie S, Tu Z. Holistically-Nested Edge Detection[J]. International Journal of Computer Vision, 2017, 125(1-3):3.

[0063] 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.

[0064] Table 1 summarizes the experimental data of Reference 1 and Example 1 on the Berkeley segmentation dataset (BSDS500). From the experimental results, the performance of Reference 1 is surpassed by using very few parameters. Figure 8 As shown in the figure, compared with the method in Reference 1, it has more refined and smoother edge information.

[0065] Table 1 Performance comparison of the edge detection method provided by the embodiment and the edge detection method of reference 1

[0066]

Claims

1. A method for edge detection of cells simulating parallel pathways XYW, characterized in that: The following steps are involved: A. Constructing a neural network, wherein the neural network includes an encoding network and a decoding network; The encoding network includes an Init module, a Path1 module, a Path2 module, a Path3 module, and a Path4 module connected in sequence; the number of channels of the Path1 module is set to C, the number of channels of the Path2 module is 2C, the number of channels of the Path3 module is 4C, and the number of channels of the Path4 module is 4C; B. The original image is input into the encoding network. After the number of channels is adjusted to C by the Init module, it is input into the Path1 module. After being processed by the Path1 module and the number of channels is adjusted to 2C, it is input into the Path2 module. After being processed by the Path2 module and the number of channels is adjusted to 4C, it is input into the Path3 module. After being processed by the Path3 module, it is input into the Path4 module. The processing result of the Path4 module is input into the decoding network. C. The decoding network decodes the processing results of the Path4 module to obtain the final detection results; The processing procedures in the Path1 module, Path2 module, and Path3 module are the same, and the processing procedures are as follows: The input results are divided into three paths. The first path is processed by three X cell modules in sequence. The number of channels is doubled during the first X cell module processing. The second path is processed by three Y cell modules in sequence. The number of channels is doubled during the first Y cell module processing. The third path is processed by three W cell modules in sequence. The number of channels is doubled during the first W cell module processing. The results of the three paths are added and fused to obtain the output result. The processing process in the Path4 module is as follows: the input result is divided into three paths, the first path is processed by three X cell modules in sequence, the second path is processed by three Y cell modules in sequence, and the third path is processed by three W cell modules in sequence; the results of the three paths are added and fused to obtain the output result; The X-cell module includes a 1×1 convolution, a 3×3 depth-separated convolution, and a subtractive fusion module. The input result is divided into two paths, one path undergoes 1×1 convolution processing, and the other path undergoes 3×3 depth-separated convolution processing. The two processing results are subtracted and fused to obtain the output result. In the first X-cell module of Path1, Path2, and Path3, the number of channels is doubled during 1×1 convolution and 3×3 depthwise separation convolution. The Y cell module includes a 1×1 convolution, a 5×5 depth-separable dilated convolution, and a subtractive fusion module. The input result is divided into two paths, one path undergoes 1×1 convolution processing, and the other path undergoes 5×5 depth-separable dilated convolution processing; the two processing results are subtracted and fused to obtain the output result; In the first Y cell module of Path1, Path2, and Path3, the number of channels is doubled during 1×1 convolution and 5×5 depthwise separable dilated convolution. The W cell module includes 1×3 depth separation convolution and 3×1 depth separation convolution. The input is processed by 1×3 depth separation convolution and 3×1 depth separation convolution in sequence to obtain the output result; In the first W cell module of the Path1 module, Path2 module, and Path3 module, the number of channels will be doubled during 1×3 depthwise separation convolution processing.

2. The edge detection method for simulating parallel path XYW cells according to claim 1, characterized in that: The Init module is a 7*7 dilated convolution with a dilation rate of 2.

3. The edge detection method for simulating parallel path XYW cells according to claim 1, wherein: The decoding network includes three ITM modules, 1×1 convolution, and Sigmoid function. The processing results of the Path4 module and the Path3 module are respectively input into the first ITM module, the processing results of the first ITM module and the Path2 module are respectively input into the second ITM module, the processing results of the second ITM module and the Path1 module are respectively input into the third ITM module, and the processing results of the third ITM module are sequentially processed by 1×1 convolution and Sigmoid function to obtain the final detection result; The processing process in the ITM module is as follows: of the two input results, the input structure with high resolution is processed by the ELC module, normalization function, and ReLU function in sequence to obtain the first processing result; the input result with low resolution is processed by the ELC module, normalization function, and ReLU function in sequence, and then upsampled to the same resolution as the high-resolution input result to obtain the second processing result; the two processing results are added to obtain the output result; The ELC module processing process is as follows: calling the convolution module of the pytorch framework, then constructing the second-order difference between the peripheral interval value and the center value of the convolution kernel, setting the intermediate convolution kernel value to zero, obtaining a new convolution module, and using the new convolution module to process the input result to obtain the output result.

Images