Dust identification method based on ultra-lightweight neural network

Abstract

The invention discloses a dust recognition method based on an ultra-light neural network, and belongs to the field of pollution prevention and control and the field of artificial intelligence. The method is realized through three steps of designing a feature extraction module, constructing a self-organizing multi-channel network structure and sharing parameters, and can achieve high-precision dust recognition performance while having ultra-light network parameters. A multi-channel structure is formed by expanding a feature extraction module in the depth and width directions, and a learnable module filtering weight and a cross-channel weight are introduced, so that the whole network structure realizes self-organization in depth and width at the same time, and all useful features can be better integrated. By utilizing a parameter sharing mechanism, the parameter quantity of the whole network model is 1k, so that the model has high execution speed, and is very suitable for industrial scenes with limited resources and high requirements on real-time performance.

Description

Technical Field

[0001] The present invention belongs to both the field of pollution prevention and control and the field of artificial intelligence. It aims to perform real-time recognition and analysis on dust images or videos through an efficient neural network model, and is applicable to scenarios such as industrial safety monitoring and environmental quality detection. Background Art

[0002] Dust recognition is of great significance in industrial production and environmental protection. Traditional methods mainly rely on physical sensors or manual detection, suffering from problems such as low detection accuracy and poor real-time performance. With the development of deep learning, image recognition technology based on convolutional neural network (CNN) has achieved remarkable results in many fields. However, in the field of dust recognition, due to the complexity and diversity of dust images, as well as the requirements for real-time performance and resource-constrained environments, traditional CNN models often fail to meet the actual application needs. In recent years, ultra-lightweight neural networks (such as MobileNet, ShuffleNet, etc.) have significantly reduced the computational amount and the number of parameters of the model while maintaining relatively high accuracy through model compression and efficient architecture design. These networks usually adopt techniques such as depthwise separable convolution, grouped convolution, and channel shuffle to improve the computational efficiency and feature extraction ability of the model. For example, MobileNet decomposes the standard convolution into depthwise convolution and pointwise convolution through depthwise separable convolution, greatly reducing the computational amount; ShuffleNet maintains the diversity of features while reducing the computational amount through grouped convolution and channel shuffle operations. Although ultra-lightweight neural networks have been widely applied in many fields, their application in the field of dust recognition is still in the exploratory stage. Dust images usually have low contrast, high noise, and complex backgrounds, which pose higher requirements for the robustness and generalization ability of the model. In addition, dust recognition needs to achieve real-time detection on resource-constrained devices, which further challenges the lightweight and efficiency of the model. Therefore, the present invention designs an ultra-lightweight neural network dedicated to dust recognition, which improves performance through a self-organizing multi-path structure and achieves ultra-lightweight of the network through parameter sharing, and can be applied to industrial environments with resource constraints and strict real-time requirements. Summary of the Invention

[0003] The present invention designs a dust recognition method based on an ultra-lightweight neural network. This method is achieved through three steps: designing a feature extraction module, constructing a self-organizing multi-path network structure, and parameter sharing, and can achieve high-precision dust recognition performance while having ultra-lightweight network parameters. The present invention shows excellent real-time performance, accuracy, and robustness in resource-constrained industrial scenarios.

[0004] The present invention is realized through the following technical solutions, including the following steps:

[0005] The first step: Design a feature extraction module;

[0006] Step 2: Construct a self-organizing multi-path network structure;

[0007] Step 3: Parameter sharing.

[0008] The creativity of the present invention is mainly reflected in:

[0009] (1) The present invention expands the feature extraction module in both the depth and width directions to form a multi-path structure, and introduces learnable module filtering weights and cross-path weights, enabling the entire network structure to achieve self-organization simultaneously in depth and width, which helps to better integrate all useful features.

[0010] (2) The present invention uses a parameter sharing mechanism, making the number of parameters of the entire network model 1k, enabling the model to have a very fast execution speed, and being very suitable for industrial scenarios with limited resources and high real-time requirements. Description of the Drawings

[0011] Figure 1 It is a flowchart of an ultra-lightweight neural network designed by the present invention for dust recognition. Specific Embodiments

[0012] The following details the embodiments of the present invention. These embodiments are implemented on the premise of the technical solution of the present invention, and detailed implementation manners and specific operation processes are given. However, the protection scope of the present invention is not limited to the following embodiments.

[0013] Embodiment:

[0014] Step 1: Design a feature extraction module;

[0015] The present invention constructs a simple lightweight feature extraction module by serially connecting a common 3×3 convolutional layer, a batch normalization (BN) layer, and a ReLU activation layer. In this module, the convolutional layer, the batch normalization layer, and the activation layer are all 1 layer.

[0016] Step 2: Construct a self-organizing multi-path network structure;

[0017] The present invention expands the feature extraction module designed in the first step in both the depth and width directions to construct a multi-path network structure for extracting features that are deep enough and shallow enough. In addition, to implement the self-organization function of the human brain, the present invention introduces learnable module filtering weights and cross-path weights, and obtains the optimal multi-path network structure by adaptively optimizing the depth, width, and weighted fusion of the outputs of different paths. Specifically, it is achieved through the following three aspects:

[0018] First, the present invention expands the feature extraction module designed in the first step in both the depth and width directions to construct a square model array of size m×m, that is, it contains m paths, and each path contains m feature extraction modules, such asFigure 1 as shown

[0019] Secondly, the present invention performs residual connection on each feature extraction module, introduces learnable module filtering weights to assign different importance to each module, and assists the self-organizing multi-path structure to achieve the optimal width and depth. The specific implementation is as follows:

[0020]

[0021] wherein, is the residual operation, and r and c are the indexes along the width and depth directions in the square model array of size m×m respectively; is the output of the feature extraction module in the r-th row and c-th column, is the output after residual connection of the feature extraction module in the r-th row and (c - 1)-th column; is the learnable module filtering weight, which can be used to adaptively adjust the depth of each path. It should be noted that, the input of is the shallow feature, denoted as F sh , and is Figure 1 the output of the 1×1 convolution in. Therefore, the output of the r-th path can be obtained through this network

[0022] Then, the present invention introduces learnable cross-path weights, and fuses the outputs of each path by weighting to obtain the depth feature F dp :

[0023]

[0024] wherein, is the learnable cross-path weight of the r-th path, which is used to adaptively adjust the width of the square model array.

[0025] Finally, the present invention passes the above depth feature F dp through a global average pooling layer (GAP) and a fully connected layer (FC) to obtain the final dust recognition result.

[0026] The third step: parameter sharing;

[0027] To make the model lighter and have a faster execution speed, inspired by the parameter sharing mechanism in multi-task learning, the present invention enforces that all feature extraction modules in the above self-organizing multi-path structure have the same weights and biases. Specifically, during the backpropagation process of training, by calculating the weights and biases of only one feature extraction module, the weights and biases of all feature extraction modules can be updated simultaneously, thus significantly improving the utilization efficiency of parameters. This step effectively avoids increasing parameters when stacking more feature extraction modules, thereby promoting a simple and efficient training and inference process, and making the dust recognition network designed by the present invention ultra-lightweight to adapt to actual industrial applications.

Claims

1. Design feature extraction module; In the present invention, a simple lightweight feature extraction module is constructed by serially connecting a common 3×3 convolutional layer, a batch normalization (BN) layer, and a ReLU activation layer. In this module, the convolutional layer, the batch normalization layer, and the activation layer are all one layer.

2. Construct a self-organizing multi-path network structure; In the present invention, the feature extraction module designed in the first step is extended simultaneously in the depth and width directions to construct a multi-path network structure for extracting features that are deep enough and shallow enough. In addition, in order to implement the self-organizing function of the human brain, the present invention introduces learnable module filter weights and cross-path weights, and obtains the optimal multi-path network structure by adaptively optimizing the depth, width, and weighted fusion of the outputs of different paths. Specifically, it is realized through the following three aspects: First, the feature extraction module designed in the first step is extended simultaneously in the depth and width directions to construct a square model array of size m×m, that is, it contains m paths, and each path contains m feature extraction modules. Second, the present invention performs residual connection on each feature extraction module, and introduces learnable module filter weights to assign different importance to each module to assist the self-organizing multi-path structure to achieve the optimal width and depth. The specific implementation is as follows: Among them, is a residual operation, where r and c are indices along the width and depth directions in an m×m sized square model array respectively; is the output of the feature extraction module at the r-th row and c-th column, is the output after residual connection by the feature extraction module at the r-th row and (c - 1)-th column; is the learnable module filtering weight, which can be used to adaptively adjust the depth of each path. It should be noted that, the input of is the shallow feature, denoted as F sh . Therefore, the output of the r-th path can be obtained through this network Then, the present invention introduces learnable cross-channel weights to weighted-fuse the outputs of each channel to obtain the depth feature F dp : Among them, is the learnable cross-channel weight of the r-th channel, which is used to adaptively adjust the width of the square model array. Finally, the present invention processes the above-mentioned depth feature F dp through a global average pooling layer (GAP) and a fully connected layer (FC) to obtain the final dust recognition result.

3. Parameter sharing; In order to make the model lighter and have a faster execution speed, inspired by the parameter sharing mechanism in multi-task learning, the present invention enforces that all feature extraction modules in the above self-organizing multi-path structure have the same weights and biases. Specifically, in the backpropagation process of training, only the weights and biases of one feature extraction module need to be calculated, and the weights and biases of all feature extraction modules can be updated simultaneously, thereby significantly improving the utilization efficiency of parameters. This step effectively avoids increasing parameters when stacking more feature extraction modules, thereby promoting the simplicity and efficiency of the training and inference processes, and making the dust recognition network designed by the present invention ultra-lightweight to adapt to actual industrial applications.

Images