Research Of Image Semantic Segmentation Algorithm Based On Multi-Scale Subtraction Network And Regional Geometric Active Contour Loss Function

Image semantic segmentation plays an extraordinary role in digital image processing and computer vision,and it is actually to assign labels to various positions in the image,and divide the pixels of the same label into the same area.Then this can be used to extract the interesting and valuable parts of the image,and provide strong support for other subsequent processing.With the rapid development of related technologies and algorithms,computer-aided medical system has become an important part of modern medical diagnosis,and image semantic segmentation has also played a huge role in the field of assisted medical treatment.Polyps are abnormal tissues in the human body,usually appearing on the mucosal surface of hollow organs,and those located in the intestinal cavity are called intestinal polyps.Generally,colorectal polyps are closely related to digestive tract malignant tumors.Once polyps become cancerous,they may lead to colorectal cancer,which is a serious harm to people’s health.At present,the identification of polyps mainly relies on the identification and judgment of colonoscopy images by professional doctors.This method mainly relies on the clinical experience of doctors and has a certain subjectivity.Moreover,it is still a tedious and complicated task to analyze and judge a large number of cases,which is easy to cause a waste of resources.Among the various methods of image segmentation nowadays,the algorithm represented by convolution neural network is favored by researchers because of its strong advantages,so many scholars have also applied relevant concepts in medical image segmentation and polyp image segmentation tasks.However,polyps tissue images have the following characteristics: 1)polyps’ size and shape are diversity and variability;2)The contrast between the real polyp and the background may not be obvious,and the boundary is fuzzy;3)The quality of the image may be poor.When the common deep learning correlation method is directly applied to the polyp image segmentation,the segmentation results may be not ideal,even wrong results.In view of this situation,based on the analysis and summary of the previous methods,this paper constructs a new algorithm for image semantic segmentation.The main contributions of this thesis is as follows:(1)In this thesis,a series of classic convolutional neural networks are analyzed and summarized to point out the shortcomings of the simple feature fusion method in the encoder-decoder structure represented by UNet.Besides,it also compares the common used loss function in the convolutional neural network.(2)A multi-scale subtraction network is proposed to improve the commonly used feature fusion operation in the form of direct contact to eliminate the redundant information between different levels of features.This model can focus on the differences between cross-level features,so as to provide more useful information for subsequent decoding operations;What is more,the regional geometric active contour model is transformed into a loss function,in which the monitoring information,such as the boundary length,interior and exterior of the polyp focus area is constructed in order to further improve the accuracy of the model segmentation of the polyp tissue image.(3)In the end,this model is trained on the public polyps image datasets,and evaluated on five commonly used polyps datasets to validate its effectiveness.Compared with the classical semantic segmentation algorithm and polyp segmentation algorithm,the proposed algorithm is proved to be available and feasible.