Research On Image Semantic Segmentation Based On Deep Learning

Image semantic segmentation plays an important role in streetscape recognition and scene understanding in autopilot systems,UAV analysis of aerial objects and object avoidance,and other consumer-grade AI terminals.As an important part of image understanding in computer vision,image semantic segmentation is not only the focus of current experts and scholars,but also an essential way for people to deepen image understanding.It is also an indispensable part of computer vision.Compared with the traditional image semantic segmentation method,no matter the cumbersome algorithm flow or the complex feature extraction method,there is no extraction method based on convolutional neural network.After analyzing and summarizing the existing semantic segmentation methods,this paper designs an image semantic segmentation network model based on cavity convolution and multi-layer feature fusion based on ResNet-101 basic network,and then combines the conditional generation confrontation network as a generation network model.The semantic segmentation network is trained by using the idea of generating confrontation to further optimize the effect of the semantic segmentation model.The main work done in this thesis is as follows:a.The development and current status of traditional image segmentation methods and image semantic segmentation based on deep learning are summarized.The related technologies involved in image semantic segmentation are described,including neural network technology,convolutional neural network technology and generation.Confrontation technology,and introduced the data sets and evaluation indicators commonly used in semantic segmentation.b.Designing a semantic segmentation based on dilation convolution and multi-layer feature fusion.ResNet-101 network is used as the encoder base network,and the convolutional convolution is used to extract the context information to maintain the feature resolution and obtain a larger local receptive field.In the decoder part,a multi-layer feature fusion method is designed to make full use of each.Hierarchical features enhance the representation of feature points,making them better for classifying image pixels.c.Applying the idea of generating anti-network to image semantic segmentation,a semantic segmentation model based on conditional generation confrontation network is designed.The model uses the semantic segmentation network model based on hole convolution and multi-layer feature fusion as the generation network,and adds the real marker map as the condition constraint.The semantic learning segmentation network is trained by the anti-learning idea to further optimize the model effect.Since the real mark map is introduced during the training process,the spatial correlation between the pixels is enhanced,so that the generated model is more accurate for the segmentation of small objects in the image.