Research On Efficient Semantic Segmentation Technology Based On Feature Enhancement

With the increasing application of vehicle and airborne imaging platforms,there is an increasing demand of intelligent imaging devices.As an important research direction of intelligent imaging devices,scene analysis that based on image segmentation technology has received more attention.This paper conducts algorithm research based on the characteristics of the image segmentation technology of the vehicle and airborne imaging platforms.The main research work is as follows:(1)The vehicle-mounted imaging platform faces complex scenes,and the lightweight segmentation algorithm does not have high segmentation accuracy on images.This paper builds a neighborhood enhanced segmentation model for these problems.This model adds a Skip decoder module and a neighborhood enhancement module to the lightweight network.It makes full use of spatial information and neighborhood correlation information without significantly increasing the amount of calculation.In addition,a phased training strategy was applied to further enhance the acquisition of multi-stage neighborhood correlation information.Experimental results show that compared with mainstream lightweight segmentation algorithms,this model has higher segmentation accuracy on vehicle images.(2)Due to the limited computing power of vehicle imaging platform,the semantic segmentation algorithm has poor real-time performance on the vehicle image.This paper builds a residual pyramid segmentation model for this problem.This model consists of a feature residual enhancement module and a Single-Shot module.The residual features are used to improve the performance of details,while the Single-Shot module is used to reduce the amount of calculation and parameters of the network.In addition,the phased training makes the network achieve better segmentation effect by gradually training the residuals at all levels.Experimental results show that compared with other algorithms with the same amount of calculations,this model has higher accuracy and better real-time performance on vehicle images.(3)The size of the target on the image obtained by the airborne imaging platform is small,and the semantic segmentation algorithm is likely to cause the small target to be mis-segmented.In response to this problem,a super-resolution segmentation model is constructed by integrating a super-resolution reconstruction algorithm.This model uses super-resolution reconstruction as a pre-module for image segmentation to supplement target details.Experimental results show that compared with the traditional upsampling method,this model has a better segmentation effect on small targets in airborne image segmentation.