Image Semantic Segmentation Based On Deep Neural Network

Image semantic segmentation based on deep neural networks plays an important role in automatic driving,autonomous navigation and other applications.In the past few years,convolutional neural networks(CNNs)have shown great potential in the image semantic segmentation task and attracted wide attention of researchers.This paper studies the applications of deep convolution neural networks in semantic segmentation of the urban road scenes.It is committed to improving the accuracy of image semantic segmentation by using binocular images and solves the problems in semantic segmentation of video sequences of the urban road scenes.The main work is as follows:(1)A novel image semantics segmentation model is proposed in the paper.It uses binocular images as input of the network and uses the depth features obtained from binocular images to compensate for the missing scene structure information due to pooling operations in traditional neural networks.In addition,in order to generate more accurate parsing maps,the idea of adversarial training is adopted in the model to further optimize the generated results of the semantic segmentation network.Experiments on the popular dataset of image semantic segmentation of the urban road scenes show that the proposed model can effectively improve the performance of traditional image semantic segmentation network,especially in some small categories,such as poles,pedestrians and so on.(2)A depth embedded recurrent predictive semantic segmentation network for video sequences is proposed in the paper.The network also uses binocular images as input,because binocular images can capture more dynamic changes between frames,thus ensuring spatiotemporal consistency of features in the sequences.At the same time,the model uses Long-short Term Memory(LSTM)to predict the features of the next frame from the continuous video sequence features,and uses the predictive features to generate the predictive semantic segmentation result.In addition,the predictive features can be combined with the features obtained from the traditional image semantic segmentation network to further improve the performance of the traditional image semantic segmentation network in the video sequence semantic segmentation task.Experiments show that our method can generate directive predictive semantic segmentation results and can effectively improve the performance of traditional image semantic segmentation models in the task of video sequence semantic segmentation.