Research On Edge Extraction Of Noisy Images Based On Deep Learning

Edge detection is an important task in computer vision,and in recent years,significant progress has been made in this task using deep learning techniques.However,challenges still exist in edge detection,such as the lack of robustness to noisy and blurry images.Most images that require edge detection usually contain some level of noise,and to achieve the best results,it is necessary to remove the noise prior to edge detection.In this thesis,a method based on a convolutional denoising autoencoder is proposed to remove noise from input images with better results compared to classical denoising methods.Furthermore,to address the problem of different importance levels of different regions in input images for edge detection,an attention mechanism is introduced in the output of each side branch of the main network to weight the feature responses of each channel,reduce attention to unimportant channels,and increase attention to important channels.This improvement method can enhance the feature representation capability of each side branch and has better effects in handling complex scenes.Through ablation experiments,this method improves ODS and OIS by 0.7% and 0.8%,respectively,compared to the original HED network.To address the issue of the limited receptive field of feature maps when the network becomes deeper,an atrous convolution method is introduced in the last stage of the HED main network,which enlarges the receptive field without increasing the computational workload.By combining atrous convolution in the last stage of the HED main network and attention mechanism in each side branch,ODS and OIS can be improved by 0.014 and 0.021,respectively.To reduce the number of parameters and computational complexity,this thesis proposes using depthwise separable convolution instead of ordinary convolution in the HED main network.The third stage uses residual connections to solve the gradient vanishing problem and improve convergence speed and accuracy.Experimental results show that using depthwise separable convolution can significantly reduce the computational complexity and model size while maintaining model accuracy.Compared to the original HED network,the number of parameters and model computational complexity are reduced by 8.3% and 5.6%,respectively,and ODS and OIS are improved by 0.1% and 0.4%,respectively.In addition,a method for improving the loss function is proposed,which introduces a tuning factor to make foreground and background easier to distinguish,resulting in a more easily convergent training process.