Research On Image Segmentation Of Street Scenes Based On Convolutional Neural Network

With the rapid development of smart devices such as mobile robots,urban drones,and self-driving cars,the safety of mobile devices of this type has become more and more of concern.Image segmentation of street scenes as a key technology to guide the movement of devices has attracted more and more attention.The more researchers are involved in this research task.The existing segmentation methods based on convolutional neural networks are generally modeled on general-purpose databases,but their performance on other databases is not ideal.The purpose of this thesis is to study image segmentation algorithms with better performance and efficiency,which can effectively avoid the deficiencies of the existing methods,and is well applicable to a specific street view database,and can be quickly operated on a mobile device.The main research contents of this thesis are as follows:1.This thesis studies an image segmentation algorithm based on feature cascading,and makes full use of the important role of each level feature in the feature cascade.A weighted combination of each channel of the feature map obtained after cascading is used to enhance the characterization capability of the feature points,so that the effect of classifying the image pixels is better.In addition,the overall segmentation performance is further improved by adding a grayscale image source at the data end.2.In this thesis,an image segmentation algorithm for small objects in streetview is studied.By extracting the features in the horizontal and vertical directions from the appropriate low-level feature map,the feature extraction of small-target objects is more adequate.And using a two-level decoder module to complete the upsampling prediction process,to further improve the effect of image prediction.3.This thesis studies a fast street-view image segmentation algorithm designed to run on the TX2 development board in real time.By adjusting the VGG-16 feature extraction network,channel-deserialization is performed on features with large resolution,channel expansion is performed on features with small resolution,and a lightweight feature extraction network is designed,which can be used on TX2.Reach the effect of rapid segmentation.The OpenCL version of the test network was implemented,further enriching the hardware use of the algorithm.Experiments show that the proposed algorithm can effectively improve the overall segmentation performance through the cascade of multi-level feature maps.Extracting features from different low-level features in different directions can effectively improve the segmentation effect on small objects.Meanwhile,a fast segmentation algorithm is designed at the mobile end.This provides practical guidance for hardware migration.