Research On Image Semantic Segmentation Algorithm Based On Convolutional Neural Network

Due to the great success of deep learning,the field of computer vision has received unprecedented attention from all walks of life.As an important part of computer vision,image semantic segmentation has been paid more and more attention by industry and academia.The convolutional neural networks have achieved unprecedented success in the field of computer vision due to their powerful feature extraction capabilities.However,the semantic segmentation network based on the fully convolutional neural network still has problems such as poor segmentation effect and high model complexity.After in-depth research on the semantic segmentation network and a summary of its shortcomings,the main work of this paper is to solve the above problems,make reasonable improvements and optimizations to the semantic segmentation network,improve the effective receptive field of the network,extract global context information,integrate multi-scale target features,improve the segmentation effect of the network for multi-scale targets,and construct an end-to-end semantic segmentation network combining global context information and multi-scale spatial pooling.In this paper,the semantic segmentation of road crack images is realized by dilated convolution and multi-channel feature fusion.Firstly,this paper improves the structure of classic encoder and decoder in the fully convolutional neural network,the high-level stage of the network is aiming at the problem of insufficient semantic information,and the large-scale feature extraction ability of the hole convolution is used to expand the receptive field of the network so that the high-level features have richer semantic information to make up for the lack of contextual information.Secondly,in the network decoder part,a multi-channel feature fusion module is added to suppress irrelevant background region responses and better restore the detailed information of the image.The crack detection model proposed in this paper has reached a 72.5% mean inetersection over union(m Io U)and96.8% F1-score on the CRACK500 and other crack detection data sets.Aiming at the problem of skin melanoma region segmentation,this paper designs a melanoma segmentation model that combines multi-scale and contextual information.In the network structure,for low-level feature output,a spatial attention mechanism is used to filter out irrelevant background information.For high-level feature output,a multi-scale receptive field is constructed,and channel attention is used to filter and weight the output feature map.The Laplace edge detection operator is used to weight the boundary pixels of the melanoma to retain more boundary information of the lesion area.Finally,the high and low layer features are weighted and fused,and the final prediction result map is output.A better segmentation effect was achieved on the ISIC2018 and PH2 skin lesion segmentation data sets.Finally,from the perspective of practical application,a semantic segmentation system for retinal blood vessel images based on a generated confrontation network is designed and implemented,which reduces the difficulty for users to train and use the image semanticsegmentation models.The generated model adopts the coding and decoding network structure and outputs segmentation results of retinal blood vessels.The discriminant model is a classical convolution neural network structure,and the output result is to classify the input image.The segmentation results of the two network models are generated and optimized through the confrontation training method of alternate iteration.The network model finally obtained is tested on the DRIVE dataset.The experimental results show that the network model based on the generated confrontation network has improved sensitivity and the segmentation results have clearer detailed performance.