Research On Image Semantic Segmentation Algorithm Based On Deep Learning

Computer vision has attracted the attention of many scholars because of its wide application scenarios.Image semantic segmentation is a difficult and practical technology in the field of computer vision.In recent years,due to the popularity of intelligent mobile terminals and the continuous development of wireless communication technologies,the explosive growth of the number of RGB images has made the research of image semantic segmentation more and more rapid.Deep learning(DL)has recently demonstrated strong performance in advanced visual tasks such as image classification and target object detection.After the introduction of fully convolutional networks(FCN),deep neural network(DNN)has made a qualitative leap in image semantic segmentation.The purpose of this paper is to design a high-performance image semantic segmentation algorithm to semantically segment images.Although most of the image semantic segmentation techniques based on deep learning methods have achieved great success,the positioning of objects in images is still Not precise enough.Therefore,this paper constructs a deep neural network model consisting of a deep convolutional neural network(DCNN)and a recurrent neural network(RNN),to some extent the existing problems.Improvement.The main work can be divided into two parts:First,a deep convolutional neural network for feature extraction is proposed.Firstly,the expansion convolution and downsampling convolution techniques are utilized in this feature extraction network to balance the conflict between feature mapping resolution and receptive field.Then,in this deep neural network,convolution,expansion convolution,downsampling convolution,batch normalization and other techniques are integrated into the residual module to form the basic unit of the network.There will be no major training errors and test errors.This feature extraction network can obtain multi-scale features and prevent loss of image detail information.Second,an end-to-end deep neural network for image semantic segmentation is constructed.Firstly,multi-scale feature fusion technology is added to the deep convolutional neural network,which enables the network to participate in multi-scale feature fusion in the process of training and prediction,and improve the accuracy of the network.Secondly,the method of transforming the fully connected condition random field into a recurrent neural network is introduced.The deep convolutional neural network and the fully connected conditional random field are integrated into an end-to-end deep neural network.This process not only makes the contour of the segmentation map more refined,but also simplifies the training process and shortens the experimental cycle.The deep neural network proposed in this paper is verified in the image semantic segmentation dataset of the public PASCAL VOC2012 dataset,and the mIOU is 78.1% in the test set.