Industrial flue gas image quality evaluation method based on digital-analog hybrid drive

Abstract

The invention discloses an industrial flue gas image quality evaluation method based on digital-analog hybrid driving, and belongs to the field of image quality evaluation. On the basis of a data-driven convolutional neural network, a self-attention mechanism and a model-driven visual saliency detection operator, the industrial flue gas image quality evaluation method based on digital-analog hybrid driving is constructed through three steps of feature extraction, feature dimension reduction and feature pooling based on digital-analog hybrid driving. According to the industrial flue gas image quality evaluation method designed by the invention, the advantages of model driving and data driving are combined, and the quality of the industrial flue gas image can be more accurately evaluated by combining a physical model, namely visual saliency detection, with a data driving algorithm, namely a self-attention mechanism. Excellent accuracy and robustness are shown in complex industrial environments and different types of flue gas emission and monitoring scenes, and a visual quality evaluation effect close to human eyes is obtained.

Description

Technical Field

[0001] The present invention belongs to the field of image quality assessment. Based on a data-driven convolutional neural network, self-attention mechanism, and a model-driven visual saliency detection operator, a method for industrial flue gas image quality assessment based on a hybrid digital-analog drive is constructed. Background Art

[0002] With the acceleration of the industrialization process, the impact of industrial flue gas emissions on the environment and human health has received increasing attention. Accurately assessing the emissions and treatment effects of industrial flue gas is of great significance for environmental protection and the sustainable development of industrial production. As an important means of monitoring and evaluating flue gas emissions, the accuracy and reliability of industrial flue gas image quality assessment methods directly affect the effectiveness of environmental monitoring and the scientific nature of decision-making.

[0003] Traditional industrial flue gas image quality assessment methods mainly rely on single model-driven or data-driven technologies. Model-driven methods usually based on physical models, such as optical models and fluid mechanics models, evaluate the quality of flue gas images by simulating the propagation and diffusion process of flue gas in the atmosphere. However, these methods often require accurate physical parameters and have poor adaptability to complex environments. Data-driven methods mainly rely on a large amount of flue gas image data and use machine learning or deep learning algorithms to automatically learn the features of flue gas images to achieve the assessment of image quality. However, data-driven methods have high requirements for the amount and quality of data, and the assessment effect will be greatly affected in the case of insufficient or poor-quality data. In recent years, hybrid digital-analog drive methods have gradually received attention. This method combines the advantages of model-driven and data-driven. By combining physical models with data-driven algorithms, it can more accurately evaluate the quality of industrial flue gas images. For example, by combining an optical model with a deep learning algorithm, it can more accurately capture the subtle changes in flue gas images, thereby improving the accuracy and reliability of the assessment.

[0004] In summary, with the continuous improvement of environmental protection requirements and the complexity of industrial production, traditional industrial flue gas image quality assessment methods are difficult to meet the actual needs. As an emerging technical means, the hybrid digital-analog drive method has high accuracy and adaptability, providing a new solution for industrial flue gas image quality assessment. Therefore, the present invention designs a method for industrial flue gas image quality assessment based on a hybrid digital-analog drive, which realizes efficient and accurate assessment of the visual quality of industrial flue gas images by combining a data-driven convolutional neural network, self-attention mechanism, and a model-driven visual saliency detection operator. Summary of the Invention

[0005] The present invention designs an industrial flue gas image quality evaluation method based on digital-analog hybrid drive. This method is achieved through three steps: feature extraction, feature dimensionality reduction, and feature pooling based on digital-analog hybrid drive, which can significantly improve the consistency between the objective evaluation results of image quality and subjective perception. The present invention shows excellent accuracy and robustness in complex industrial environments, different types of flue gas emissions, and monitoring scenarios, achieving a visual quality evaluation effect close to that of the human eye.

[0006] The present invention is achieved through the following technical solutions, including the following steps:

[0007] The first step: Feature extraction;

[0008] The second step: Feature dimensionality reduction;

[0009] The third step: Feature pooling based on digital-analog hybrid drive.

[0010] The creativity of the present invention is mainly reflected in:

[0011] The industrial flue gas image quality evaluation method designed by the present invention combines the advantages of model drive and data drive. By combining the physical model of visual saliency detection with the data-driven algorithm of self-attention mechanism, it can more accurately evaluate the quality of industrial flue gas images. Description of the Drawings

[0012] Figure 1 is the network framework diagram of the industrial flue gas image quality evaluation based on digital-analog hybrid drive designed by the present invention. Detailed Embodiments

[0013] The following details the embodiments of the present invention. These embodiments are implemented on the premise of the technical solutions 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.

[0014] Embodiment:

[0015] The first step: Feature extraction;

[0016] To make the image quality evaluation more efficient and rapid, the present invention designs an ultra-lightweight feature extraction module. This module is serially connected by 1 simple 5×5 convolutional layer, 1 batch normalization layer, and 1 ReLU activation layer. At the same time, the number of channels is set to 2. Compared with ordinary convolutional neural network modules with hundreds or thousands of channels, the parameter quantity of the feature extraction module designed by the present invention is significantly reduced. This feature extraction module accepts the input reference image and distorted image in a parameter-sharing manner, and combines the extracted reference image features and distorted image features in parallel to generate feature F, as Figure 1 shown.

[0017] The second step: Feature dimensionality reduction;

[0018] First, the present invention reduces the channel dimension of the feature F to 1 through a multi-layer perceptron (MLP) to obtain a feature Subsequently, an average pooling layer with a stride of 16×16 is used to reduce the feature in both the width and height dimensions to 1 / 16 of the input image size, generating a feature to further reduce the subsequent computational complexity and accelerate the model calculation speed.

[0019] Step 3: Feature pooling based on digital-analog hybrid drive;

[0020] To better perform feature pooling, the present invention designs a new feature pooling module based on digital-analog hybrid drive.

[0021] First, the present invention uses a data-driven self-attention mechanism to obtain the attention matrix M of the feature , which can highlight the important regions in the feature A and suppress the unimportant regions. The self-attention mechanism has the advantage of capturing the long-range dependence relationships in the feature , enabling the attention matrix M to better reflect the global information of the feature A .

[0022] Secondly, the present invention uses a model-driven visual saliency detection operator to calculate the salient regions in the input image, obtaining a saliency map M S , and these saliency maps M S are usually related to the target objects or important features in the input image and can provide more accurate local information.

[0023] Then, the present invention performs a dimensionality transformation on the saliency map M S to generate to align its dimension with that of the attention matrix M A , and sets two weights (ω1 and ω2) to automatically adjust the importance of and M A to generate a new attention matrix M. Inspired by the fact that residual connections can improve the model performance, the present invention adds a diagonal matrix I with the same dimension as M to M to further improve the representation ability of feature pooling and enhance the model's ability to capture key information. Thus, the final digital-analog hybrid drive attention matrix M DM is obtained.

[0024] Finally, the above digital-analog hybrid drive attention matrix M DM is used to process the feature Perform weighted pooling to obtain the visual perception quality score of the industrial flue gas image. The implementation formula is as follows:

[0025]

[0026] where score is the visual perception quality score of the industrial flue gas image; is the matrix multiplication operation.

Claims

1. Feature extraction: In order to make image quality evaluation more efficient and fast, the present invention designs an ultra-light feature extraction module. The module is composed of a simple 5×5 convolution layer, a batch normalization layer and a ReLU activation layer connected in series, and the number of channels is set to 2. Compared with ordinary convolutional neural network modules with hundreds or thousands of channels, the number of parameters of the feature extraction module designed by the present invention is significantly reduced. The feature extraction module accepts the input reference image and distorted image in a parameter sharing manner, and extracts the reference image features and the distorted image features in parallel to generate feature F.

2. Feature Dimensionality Reduction: First, the present invention reduces the channel dimension of the feature F to 1 through a multi-layer perceptron (MLP) to obtain a feature Subsequently, an average pooling layer with a stride of 16×16 is used to reduce the feature in both the width and height dimensions to 1 / 16 of the input image size, generating a feature to further reduce the subsequent computational complexity and accelerate the model calculation speed.

3. Feature pooling based on digital-analog hybrid drive: In order to better perform feature pooling, the present invention designs a new feature pooling module based on digital-analog hybrid drive. First, the present invention utilizes a data-driven self-attention mechanism to obtain the feature attention matrix M A , which can highlight the important regions in the feature and suppress the unimportant regions. The self-attention mechanism has the advantage of capturing the long-range dependence relationships in the feature , enabling the attention matrix M A to better reflect the global information of the feature . Secondly, the present invention uses a model-driven visual saliency detection operator to calculate the salient regions in the input image and obtain a saliency map M S , and these saliency maps M S are usually related to the target objects or important features in the input image and can provide more accurate local information. Then, the present invention performs dimensionality transformation on the saliency map M S to generate so that its dimension is aligned with that of the attention matrix M A , and two weights (ω1 and ω2) are set to automatically adjust and M A importance, to generate a new attention matrix M. Inspired by the fact that residual connections can improve the performance of the model, the present invention adds a diagonal matrix I with the same dimension as M to M, further improving the representation ability of feature pooling and enhancing the model's ability to capture key information. Thus, the final digital-analog hybrid-driven attention matrix M DM is obtained. Finally, the above digital-analog hybrid-driven attention matrix M is used DM to weight pooling on the features to obtain the visual perception quality score of the industrial flue gas image. The implementation formula is as follows: where score is the visual perception quality score of the industrial flue gas image; is a matrix multiplication operation.

Images