Research On Semantic Segmentation Of Multi-scale Complex Scene Images Based On Dual Channels

The rapid development of deep learning and the massive increase of image data have promoted the application of image semantic segmentation technology.Simultaneously,with intelligent development of society,image semantic segmentation has a wide range of applications in diverse areas such as autonomous driving,facial segmentation,fashion classification,and agricultural intelligence.Due to the superiority of deep learning in processing massive amounts of data and the natural advantages of image processing,convolutional neural networks based on deep learning have become the mainstream.Image semantic segmentation methods based on convolutional neural networks are divided into image feature extraction stage,image feature fusion stage and semantic segmentation image prediction stage.Currently,the mainstream direction of researches is to find more reliable feature extraction modules and better segmentation models for image segmentation under complex scenes.This paper focuses on the image size interence,the excessive network parameters,the insufficient utilization of image feature information,and the flawed fusion ways,etc.Thus,this paper propose a multi-level fusion pyramid neural network model and a dual-channel multi-scale feature neural network model.The main work of the paper is as follows:Firstly,the common image features and the principle of convolutional neural network are analyzed.This paper establish the features,basic methods and evaluation indexes which are used in image semantic segmentation.Additionally,appropriate image preprocessings are conducted according to the different scenes from different datasets.Secondly,for the problem of both the rough feature processings and too many parameters,which are resulted by the traditional full convolutional neural network.In this paper,the feature extraction block and pyramid module are improved from the perspective of image feature extraction and feature fusion.By optimizing the feature extraction module,the number of parameters of the feature extraction module is reduced by 9.43% without the degradation of the network performance.In addition,the dilated convolution module is used to improve the pyramid module,which optimizes the original feature extraction and fusion method to make it more suitable for complex image scenarios.Thus,a network model based on a multi-level fusion pyramid was constructed in this paper.Through ablation experiments and contrast experiments,the feasibility of the improved method and model was verified.Compared with other feature extraction models,the performance is improved by 1.8%.Finally,an adaptive weight block and a multi-level feature fusion block are proposed to enhance the feature fusion,making the feature image can contain more details and spatial information of the image,and alleviating the feature loss after continuous feature extraction.Based on the shortcomings of low-level utilization of image feature information,this paper proposes a local feature processing channel and a global feature correlation channel,which based on a multi-level feature fusion module to solve it,and builds a multi-scale feature neural network model based on dual channels.The optimal parameter settings of the dual-channel model are determined through comparsion experiments.The model in this paper performs excellently in image segmentation in multiple scenes.Compared with other models,the performance is improved by 1.39%to 1.91%,and it can complete the task of semantic segmentation of complex scene images.