Research On U-Net Neural Network With Half-Channel Attention Mechanism For Industrial Image Restoration

Industrial image restoration has attracted a lot of attention in the field of image processing and image restoration,and industrial image restoration has an important significance.However,there are still some problems in the research of image restoration techniques in the industrial field.For example,low accuracy,large number of model parameters and difficulty in balancing the performance and size of models.Therefore,in order to solve the above-mentioned defects,three industrial image restoration models are proposed,including a two-stage AHINet fusing half-channel attention mechanism,a multi-stage MSFFNet fusing attention features,and a DAResNet fusing half-channel attention mechanism and double-layer residual blocks.The main research work is summarized as follows:(1)To address the problem of low accuracy of the fuzzy images of exhaust gas discs collected by industry,a targeted half-channel attention mechanism is designed by analyzing the image features,constructing a half-instance normalization block based on the half-channel attention mechanism,and proposing a two-stage AHINet neural network incorporating the half-channel attention mechanism.The half-channel attention mechanism in this model removes the maximum pooling layer from the traditional channel attention mechanism and uses only the average pooling layer for the extraction of low and medium-level features,which makes the half-instance normalization block of the half-channel attention mechanism more targeted for the collected blurred images of industrial exhaust gas discs and ensures the maximum retention of image features without adding additional parameters.In addition,we produced a dataset of industrial exhaust gas disk fuzzy images.A comparison experiment was conducted with HINet neural network on this dataset,and the final experimental results showed that the model achieved better results in both evaluation metrics.(2)Although the two-stage AHINet neural network fused with half-channel attention mechanism achieves good results in processing industrial images,the number of neural network parameters and the network model are too large.Therefore,based on this situation,this thesis designs a multi-stage neural network model to further compress the model size of each stage,and adds a half-instance normalization block based on the half-channel attention mechanism to propose a multi-stage MSFFNet neural network with fused attention features.A multi-stage fusion module is also designed to enable further enhancement of image features and contextual feature fusion.Finally,the proposed MSFFNet neural network is compared with HINet and AHINet in experiments,in which the model size of the three neural networks is focused on comparing,and the final experimental results show that the model guarantees a maximum reduction of 72.6% in the number of parameters with a minimum decrease of 0.0475 dB in accuracy.(3)To address the problem of balancing the number of neural network model parameters with image restoration accuracy,a DAResNet neural network incorporating a half-channel attention mechanism and a double-layer residual block is constructed to reduce the number of neural network parameters while ensuring accuracy.The model is composed of a double-layer residual block and a half-instance normalization block based on the half-channel attention mechanism.The half-channel attention mechanism half-instance normalization block is used to extract features at each scale to enhance low-level features and mid-level features of the image,and a double-layer residual block is used to enhance the extraction of high-level features.Meanwhile,we select four recently proposed excellent algorithms on the proposed dataset for comparison,and DAResNet improves the image restoration accuracy by 0.1111 dB while the number of parameters is only 6.56%,and is validated by ablation experiments to demonstrate the excellence and compatibility of each of the proposed blocks.