Research On Image Semantic Segmentation Methods Based On CNN

Image semantic segmentation refers to the process of separating the foreground and background of an image according to certain criteria and assigning predefined category labels to the pixels of each foreground target.Image semantic segmentation is an cross field of computer vision,image processing,and artificial intelligence,which is widely used in medical image analysis,automatic driving and other fields.With the further development of artificial intelligence theories,many researchers have improved convolutional neural networks(CNN)and applied them to segmentation tasks,and made many breakthroughs.The segmentation model Dense ASPP proposed in 2018 is based on the convolutional neural network Dense Net,and its DASPP module is improved from Dense Net's dense block.Dense ASPP has high segmentation accuracy,but it has a large amount of network parameters and high complexity,and cannot fully extract the context information of the image.Aiming at the shortcomings of Dense ASPP,this paper studies the image semantic segmentation method based on convolutional neural network.The main research contents of this paper include:1.Aiming at the problems of high complexity and poor real-time performance of Dense Net,which is the backbone network of Dense ASPP model,an image semantic segmentation method based on Shufflenetv2 is proposed.The lightweight convolutional neural network Shufflenetv2 is used instead of Dense Net as the backbone network of the segmentation model to extract features,which effectively reduces the amount of parameters and calculation of the model and improves the real-time performance of the segmentation algorithm.2.Aiming at the problems of limited global information obtained by the atrous convolutions and different importance of each output feature in the DASPP module,a method for improving DASPP by combining global average pooling and the SE module is proposed.In DASPP module,a global average pooling layer is added in parallel to obtain global information,and then a SE module is added to balance the importance of each channel to achieve image semantic segmentation.The experiment results show that this method improves the accuracy of image semantic segmentation.3.Aiming at the problem that the upsampling part of the model is too rough,an image semantic segmentation method based on encoding-decoding structure is proposed.This method up-samples twice in the decoding stage,and after the up-sampling,merging it with the feature map of the same resolution in the lower layer,adds detailed information to get a more effective decoder.The results of experiment show that this method refines the segmentation results,improves the segmentation accuracy of the model,and reflects the feasibility.