Research On Image Semantic Segmentation Algorithm Based On Deep Learning

Image semantic segmentation is an important step in making computers understand complex scenes like humans,it is also one of the research hotspots in the field of computer vision.This task is the process of assigning labels to each pixel in an image for the convenience of image analysis.With the continuous breakthrough of deep learning in the field of computer vision,image semantic segmentation has also received more and more attention from computer vision and machine learning researchers.The research work of this article mainly includes the following two aspects:(1)Aiming at the problem that the fully convolutional network is easy to lose detailed information during semantic segmentation,resulting in rough segmentation results,and the large amount of calculations in the feature extraction network,an improved algorithm based on the fully convolutional network is proposed,mainly in the feature extraction and upsampling process has been improved.Using several depthwise separable convolutional blocks for feature extraction,compared with the feature extraction network of the fully convolution network,the model parameters and the amount of calculation are reduced;The fully convolution network directly increases the size of feature map through deconvolution in the upsampling process,and loses a lot of details information.In response to this problem,we used the pyramid pooling module in the upsampling process,and at the same time,the feature maps obtained during the upsampling process and the feature maps of the same size during the feature extraction process were spliced together in the channel dimension to achieve multi-scale feature fusion.Therefore,the improved network can aggregate more contextual information,effectively capture rich multi-scale information,and reduce information loss.Experimental results on PASCAL VOC dataset and Cityscapes dataset show that this method has better segmentation effect compared with other deep learning algorithms.(2)The precise segmentation of retinal blood vessels is of great guiding significance for the early diagnosis of some diseases.We propose a retinal blood vessel image segmentation algorithm based on encoder-decoder structure.In the encoder stage,the Inception module is used to perform feature extraction on the image,and the multi-scale information of the image can be obtained by using different scale convolution kernels;In order to enable the network to perceive small blood vessels of different scales,and improve the accuracy of small blood vessel segmentation,in the decoder stage,multiple pyramid pooling modules are used to aggregate more contextual information,multi-scale local area feature fusion can improve the effect of small blood vessel segmentation;In addition,in the process of upsampling,the feature fusion method is used to fuse low-level semantic features to obtain more low-level detailed information,and further improve the segmentation accuracy of retinal blood vessel images.Experiments on the DRIVE and STAER fundus image datasets verify the effectiveness of the method.