Research On End-to-End Image Semantic Segmentation Algorithm

With the rapid development of artificial intelligence and deep learning,image semantic segmentation has gradually become a hot issue in the field of computer vision research,and it is also gradually applied to the field of medical image segmentation.However,due to the continuous complexity of the segmentation model,the existing algorithms take up too much computing resources and make the model parameters redundant.At the same time,the existing semantic segmentation algorithms have not achieved good results in some medical image segmentation tasks because the size and shape of segmentation objects vary,and the distribution of positive and negative samples is unbalanced.The above reasons make designing an efficient and robust semantic segmentation algorithm still facing huge challenges.In this paper,the semantic segmentation algorithm is studied deeply,and the existing segmentation method is improved to ameliorate the accuracy of segmentation.The main work of the article is as follows:Firstly,the domestic and foreign research status of semantic segmentation technology is analyzed,the difficulties faced in the field of medical image segmentation are analyzed,the basic principle of image semantic segmentation algorithm is explained,and the two mainstream algorithms of semantic segmentation are elaborated,the advantages and disadvantages of mainstream semantic segmentation network structures such as full convolutional neural network,encoder-decoder structure and expanded convolution are analyzed.Secondly,in view of the problem of excessive and redundant use of computing resources and model parameters caused by the current semantic segmentation network relying on multi-level cascaded convolutional neural network to extract the region of interest,an improved U-Net semantic segmentation based on attention mechanism is proposed In the feature extraction method,the attention gating module is incorporated into the standard U-Net model,focusing on displaying the advanced features passed by skipping the connection structure to avoid repeated extraction of extraneous features.Then the method flow and the improved U-Net network structure are described in detail,and the superiority of the improved scheme is proved through experimental comparison.Then,in response to the problem of the reduction of batch normalization statistical accuracy faced by the semantic segmentation network during the training of small batches,a deeplab v3+ algorithm based on group normalization is proposed.Based on the original network model,the original batch The sub-normalization layer is adjusted to a group normalization layer,which effectively balances the contradiction between the memory resources occupied by training and the calculation error,and the Lovasz-softmax loss function is added to adapt to the semantic segmentation problem in more scenarios.Finally,in view of the imbalance of label data categories in semantic segmentation,the sensitivity of common loss functions to the problem of data category imbalance is analyzed.The generalized Dice loss function is selected and integrated into the existing u-net algorithm.Finally,the superiority and robustness of this method are proved through comparative experiments.