Research And Implementation Of Image Semantic Segmentation Based On Deep Learning

In recent years,with continuous innovation and development of deep learning theory,image recognition,image detection and image semantic segmentation technologies based on the deep learning are gradually replacing the traditional artificially-designed scheme.Image semantic segmentation is an important research direction in computer vision,it is also an important part in image understanding.The research of image semantic segmentation is one of the hottest topics in academic world,as well as in industrial demand.Within the progress of human civilization,intelligent driving,indoor navigation,and human-machine interaction require precise,efficient segmentation mechanisms,therefore,the study of the semantic segmentation technique has important practical significance.As one of the methods of deep learning,convolutional neural network has strong feature extraction capability and has achieved unprecedented success in computer vision field.After studying FCN(Fully Convolutional Networks)and summarizing its deficiencies,this thesis redesigned and implemented an end-to-end image semantic segmentation network model.The specific results are as follows:First of all,the newly designed network model which added the global average pooling layer,the context information of the image is merged,the ability of network model to deal with image details is improved;The L2 normalization layer added with training parameters reduces the difficulty of network training and improves the network convergence speed;Using the PReLU activation function with the training parameters,the nonlinear modeling ability of the network is improved,the stochastic gradient descent is closer to the natural gradient,and the network model converges faster.Secondly,this thesis designed training and testing methods of the newly designed network model,trained and tested the newly designed network model on the PASCAL VOC 2012 datasets,and then optimized the network parameters,analyzed and summarized the experimental data through the network model information visualization and the comparison with other semantic segmentation models.The experimental results indicate that although newly designed network model added multiple network layers and additional training parameters relative to FCN,the benefits outweigh the disadvantages: Overall training time of newly designed network model is less than half of fcn-8s' s,the segmentation effect is better,and MIoU increases by 4.6 percentage points.Finally,this thesis designed and implemented an image semantic segmentation system for practical application based on the newly designed network model,which reduced the difficulty for users to train and use image semantic segmentation models.