A lightweight contour detection method simulating the parallel hierarchical processing mechanism of a vision system
By simulating the parallel hierarchical processing mechanism of the visual system, a lightweight neural network structure is constructed, which solves the problem of high computational cost of convolutional neural networks in scenarios with scarce computing resources, and achieves high-efficiency contour detection performance.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-07-25
- Publication Date
- 2026-04-03
AI Technical Summary
Existing contour detection models based on convolutional neural networks have high computational requirements and a large number of parameters, making them difficult to run efficiently in application scenarios with limited computing resources.
A lightweight neural network structure is constructed by simulating the parallel hierarchical processing mechanism of a vision system, including X-shaped, Y-shaped, and simple and complex cell models. Combined with direction-selective convolution and feature integration modules, a lightweight contour detection neural network is designed.
With minimal parameters, contour detection performance comparable to traditional methods was achieved, improving computational and processing efficiency.
Smart Images
Figure CN115272672B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image processing, and more specifically to a lightweight contour detection method that simulates the parallel hierarchical processing mechanism of a visual system. Background Technology
[0002] Contour detection aims to extract the boundary between the background and the target in an image. It is often used as a key step in the front-end processing of various intermediate and advanced computer vision tasks and is one of the fundamental tasks in the field of computer vision research.
[0003] Contour detection aims to extract meaningful contours from natural images, thereby assisting in the processing of some high-level visual tasks; it is a relatively low-level visual task. Thanks to the development of deep learning technology, contour detection models based on convolutional neural networks have demonstrated strong performance. These methods use transfer learning to pre-train classic networks such as VGG or ResNet to obtain encoding networks, and then design decoding networks to form end-to-end contour detection models. Experiments on the Berkeley Segmentation Dataset (BSDS500) have demonstrated the performance of these models.
[0004] While current contour detection models based on convolutional neural networks have achieved good performance, they employ classic networks such as VGG, resulting in a large number of training parameters and high computational cost, making them unsuitable for some application scenarios with limited computing resources. Summary of the Invention
[0005] This invention aims to provide a lightweight contour detection method that simulates the parallel hierarchical processing mechanism of the visual system. Starting from enhancing the computational efficiency of the deep learning model for contour detection, this method designs a lightweight contour detection neural network based on the biological visual mechanism. This network can achieve performance comparable to conventional encoder-decoder contour detection networks with very few parameters.
[0006] The technical solution of the present invention is as follows:
[0007] The lightweight contour detection method based on the parallel hierarchical processing mechanism of the simulated vision system includes the following steps:
[0008] A. Construct the neural network structure, as follows:
[0009] Encoding networks, decoding networks;
[0010] The encoding network includes X-type cell model, Y-type cell model, simple cell model, and complex cell model; the simple cell model is simulated by direction-selective convolution, and the complex cell model is simulated by the output of the simple cell model again through direction-selective convolution.
[0011] The decoding network consists of a feature integration module;
[0012] B. The original image is input into the neural network. First, it is integrated through 1*1 convolution of each channel and then input into the X-type cell model and Y-type cell model respectively. The outputs of the X-type cell model and Y-type cell model are integrated through 1*1 convolution of each channel respectively and then fused together. The fused result is input into the simple cell model and the decoding network respectively. The output of the simple cell model is integrated through 1*1 convolution of each channel and then input into the complex cell model and the decoding network respectively. The output of the complex cell model is integrated through 1*1 convolution of each channel and then input into the decoding network. The decoding network decodes and outputs the final contour.
[0013] The X-type cell model is as follows:
[0014] (1)
[0015] (2)
[0016] (3)
[0017] (4)
[0018] (5)
[0019] in, Channel number; The learnable weight parameters control the intensity of the inhibition of the central mechanism by the peripheral mechanism; and These are the central mechanism and the peripheral mechanism in the 1st... i Learnable convolutional kernels for each channel; It is a radius of The circle; and The inner radius is outer radius is The ring; and These are functions for calculating the mean and variance, respectively. The i-th channel represents the input feature, A represents a placeholder, and represents the independent variable of this function.
[0020] The Y-type cell model is as follows:
[0021] (6)
[0022] (7)
[0023] (8)
[0024] in, For channel labeling, The intensity of inhibition of the central mechanism by the peripheral mechanisms was controlled, which was obtained through learning; and These are the central mechanism response and the peripheral mechanism response; The global response for all nonlinear subunits is obtained by averaging the responses of all nonlinear subunits. k Represents the number of nonlinear subunits; and The first j Learnable convolutional kernels of nonlinear subunits and their responses when acting alone; Responsible for the first j The responses of each nonlinear subunit in its assigned region are collected at the central location. The central location has the same shape and size as the region handled by the nonlinear subunit, being a ring with a width of [missing information]. w ; The i-th channel represents the input feature.
[0025] The directional selective convolution is:
[0026] (9)
[0027] in, This is a 2×2 pooling operation. It is a learnable convolutional kernel in the j-th direction (out of 4 directions).
[0028] The decoding network consists of two feature integration modules. The output results of the simple cell model and the complex cell model are integrated by performing 1*1 convolution on each channel and then input into the first feature integration module. The first feature integration module processes the data to obtain the first integrated feature. The first integrated feature and the result of addition and fusion are input into the second feature integration module. The second feature integration module processes the data to obtain the second integrated feature, which is the final contour.
[0029] In the feature integration module, low-resolution features in the input features are processed by 3*3 convolution, then instance normalization, then ReLU function processing, and finally permutation convolution to obtain the permutation result; high-resolution features in the input features are processed by 3*3 convolution, then instance normalization, then ReLU function processing, and finally added to the permutation result to obtain the integrated features.
[0030] The method of the present invention can achieve performance comparable to that of conventional methods with very few parameters.
[0031] Based on biological vision mechanisms, this invention establishes a convolutional model that differs from conventional methods, making neural networks more efficient in feature extraction and processing, and demonstrating high processing efficiency in experiments. Attached Figure Description
[0032] Figure 1 This is a schematic diagram of the overall structure of the neural network in Embodiment 1 of the present invention;
[0033] Figure 2 This is a schematic diagram of the decoding network structure in Embodiment 1 of the present invention;
[0034] Figure 3 This is a schematic diagram of the integrated feature module of Embodiment 1 of the present invention. Detailed Implementation
[0035] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. Example 1
[0036] This embodiment provides a lightweight contour detection method that simulates the parallel hierarchical processing mechanism of a vision system, including the following steps:
[0037] A. Construct the neural network structure, as follows:
[0038] Encoding networks, decoding networks;
[0039] The encoding network includes X-type cell model, Y-type cell model, simple cell model, and complex cell model; the simple cell model is simulated by direction-selective convolution, and the complex cell model is simulated by the output of the simple cell model again through direction-selective convolution.
[0040] The decoding network consists of a feature integration module;
[0041] B. The original image is input into the neural network. First, it is integrated through 1*1 convolution of each channel and then input into the X-type cell model and Y-type cell model respectively. The outputs of the X-type cell model and Y-type cell model are integrated through 1*1 convolution of each channel respectively and then fused together. The fused result is input into the simple cell model and the decoding network respectively. The output of the simple cell model is integrated through 1*1 convolution of each channel and then input into the complex cell model and the decoding network respectively. The output of the complex cell model is integrated through 1*1 convolution of each channel and then input into the decoding network. The decoding network decodes and outputs the final contour.
[0042] The X-type cell model is as follows:
[0043] (1)
[0044] (2)
[0045] (3)
[0046] (4)
[0047] (5)
[0048] in, Channel number; The learnable weight parameters control the intensity of the inhibition of the central mechanism by the peripheral mechanism; and These are the central mechanism and the peripheral mechanism in the 1st... i Learnable convolutional kernels for each channel; It is a radius of (In this embodiment, it is set to a circle of 5); while The inner radius is outer radius is (In this embodiment, it is set to 21) rings; and These are functions for calculating the mean and variance, respectively. The i-th channel represents the input feature, A represents a placeholder, and represents the independent variable of this function.
[0049] The Y-type cell model is as follows:
[0050] (6)
[0051] (7)
[0052] (8)
[0053] in, For channel labeling, The intensity of inhibition of the central mechanism by the peripheral mechanisms was controlled, which was obtained through learning; and These are the central mechanism response and the peripheral mechanism response; The global response for all nonlinear subunits is obtained by averaging the responses of all nonlinear subunits. k Represents the number of nonlinear subunits; and The first jLearnable convolutional kernels of nonlinear subunits and their responses when acting alone; Responsible for the first j The responses of each nonlinear subunit in its assigned region are collected at the central location. The central location has the same shape and size as the region handled by the nonlinear subunit, being a ring with a width of [missing information]. w ; The i-th channel represents the input feature.
[0054] The directional selective convolution is:
[0055] (9)
[0056] in, This is a 2×2 pooling operation. It is a learnable convolutional kernel in the j-th direction (out of 4 directions).
[0057] The decoding network consists of two feature integration modules. The output results of the simple cell model and the complex cell model are integrated by performing 1*1 convolution on each channel and then input into the first feature integration module. The first feature integration module processes the data to obtain the first integrated feature. The first integrated feature and the result of addition and fusion are input into the second feature integration module. The second feature integration module processes the data to obtain the second integrated feature, which is the final contour.
[0058] In the feature integration module, low-resolution features in the input features are processed by 3*3 convolution, then instance normalization, then ReLU function processing, and finally permutation convolution to obtain the permutation result; high-resolution features in the input features are processed by 3*3 convolution, then instance normalization, then ReLU function processing, and finally added to the permutation result to obtain the integrated features. Example 2
[0059] For quantitative performance evaluation of the final contour plot, in addition to using the same performance measurement standard ODS as in Reference 1, we also used the parameter quantity Params and floating-point computational quantity FLOPs. A higher ODS indicates better model accuracy. The ODS calculation method is as follows:
[0060] (10)
[0061] (11)
[0062] (12)
[0063] in This indicates the correctly detected contour pixels. This indicates the number of non-contour pixels that were detected as contour pixels. This represents the number of contour pixels detected as non-contour pixels. Params is the number of learnable parameters in the model, and FLOPs is the number of floating-point operations performed by the model when processing data. For convolutional layers without bias terms, the calculation method is as follows:
[0064] (13)
[0065] in C i Input the number of channels. H and W For the height and width of the input feature, C 0 represents the number of output channels.
[0066] Document 1: He J, Zhang S, Yang M, et al. Bi-directional cascade network for perceptual edge detection[C]. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Long Beach, 2019:3828-3837.
[0067] The parameters used in the evaluation of Reference 1 and Example 1 are the same.
[0068] Table 1 shows the evaluation results of Reference 1 and Example 1 on the Berkeley Segmentation Dataset (BSDS500). Example 1 has fewer parameters and fewer FLOPs compared to Reference 1, but has a higher ODS.
[0069] Model Name source ODS Params (M) FLOPs (G) BDCN2 Reference 1 0.766 0.28 22.6 XYS Example 1 0.770 0.14 11.9 .
Claims
1. A lightweight contour detection method that simulates the parallel hierarchical processing mechanism of a visual system, characterized in that... Includes the following steps: A. Construct the neural network structure, as follows: Encoding networks, decoding networks; The encoding network includes X-type cell model, Y-type cell model, simple cell model, and complex cell model; the simple cell model is simulated by direction-selective convolution, and the complex cell model is simulated by the output of the simple cell model again through direction-selective convolution. The decoding network consists of a feature integration module; B. The original image is input into the neural network. First, it is integrated through 1*1 convolution of each channel and then input into the X-type cell model and Y-type cell model respectively. The outputs of the X-type cell model and Y-type cell model are integrated through 1*1 convolution of each channel respectively and then fused together. The fused result is input into the simple cell model and the decoding network respectively. The output of the simple cell model is integrated through 1*1 convolution of each channel and then input into the complex cell model and the decoding network respectively. The output of the complex cell model is integrated through 1*1 convolution of each channel and then input into the decoding network. The decoding network decodes and outputs the final contour.
2. The lightweight contour detection method based on the parallel hierarchical processing mechanism of a simulated vision system as described in claim 1, characterized in that: The X-type cell model is as follows: （1） （2） （3） （4） （5） in, Channel number; The learnable weight parameters control the intensity of the inhibition of the central mechanism by the peripheral mechanism; and These are the central mechanism and the peripheral mechanism in the 1st... i Learnable convolutional kernels for each channel; It is a radius of The circle; and The inner radius is outer radius is The ring; and These are functions for calculating the mean and variance, respectively. The i-th channel represents the input feature; A represents a placeholder, which is the independent variable of this function.
3. The lightweight contour detection method based on the parallel hierarchical processing mechanism of a simulated vision system as described in claim 1, characterized in that: The Y-type cell model is as follows: (6) （2） （3） （7） （8） in, For channel labeling, The intensity of inhibition of the central mechanism by the peripheral mechanisms was controlled, which was obtained through learning; and These are the central mechanism response and the peripheral mechanism response; The global response for all nonlinear subunits is obtained by averaging the responses of all nonlinear subunits. k Represents the number of nonlinear subunits; and These are the central mechanism and the peripheral mechanism in the 1st... i Learnable convolutional kernels for each channel; It is a radius of The circle; and The inner radius is outer radius is The ring; and The first j Learnable convolutional kernels of nonlinear subunits and their responses when acting alone; Responsible for the first j The responses of each nonlinear subunit in its assigned region are collected at the central location. The central location has the same shape and size as the region handled by the nonlinear subunit, being a ring with a width of [missing information]. w ; The i-th channel represents the input feature.
4. The lightweight contour detection method based on the parallel hierarchical processing mechanism of a simulated vision system as described in claim 1, characterized in that: The directional selective convolution is: （9） in, This is a 2×2 pooling operation; There are 4 learnable convolutional kernels in the j-th direction.
5. The lightweight contour detection method based on the parallel hierarchical processing mechanism of a simulated vision system as described in claim 1, characterized in that: The decoding network consists of two feature integration modules. The output results of the simple cell model and the complex cell model are integrated by performing 1*1 convolution on each channel and then input into the first feature integration module. The first feature integration module processes the data to obtain the first integrated feature. The first integrated feature and the result of addition and fusion are input into the second feature integration module. The second feature integration module processes the data to obtain the second integrated feature, which is the final contour.
6. The lightweight contour detection method based on the parallel hierarchical processing mechanism of a simulated vision system as described in claim 5, characterized in that: In the feature integration module, low-resolution features in the input features are processed by 3*3 convolution, then instance normalization, then ReLU function processing, and finally permutation convolution to obtain the permutation result; high-resolution features in the input features are processed by 3*3 convolution, then instance normalization, then ReLU function processing, and finally added to the permutation result to obtain the integrated features.