Research And Implementation Of Image Semantic Segmentation Algorithm Based On Deep Learning

Image semantic segmentation task is one of the research hotspots in the field of computer vision.Each pixel of the images is classified in this task.With the development of smart devices such as self-driving cars and drones,accurate extraction of image information is the primary question that must be considered by researchers.Image semantic segmentation as a key technology to solve this problem has attracted more and more researchers who devote themselves to this field.This thesis aims to study the high-performance lightweight street view image semantic segmentation algorithm.First,the existing image semantic segmentation algorithm is analyzed in this thesis.Based on this algorithm,it can be applied to the existing street view data set by optimization algorithm.And then,the algorithm has been further improved,so the fast street view segmentation algorithm is implemented on the mobile terminal.The content of this thesis is as follows:1.This thesis studies a lightweight street scene image semantic segmentation algorithm based on attention alignment.The existing basic network consumes a lot of computing resources and does not meet the requirements of light weight.In this thesis,lightweight and efficient DF1 is used as the basic network to reduce the amount of computing while meeting the accuracy requirements.In view of the problem that the existing ASPP module has too much calculation and lack of correlation between branches,this thesis proposes a horizontally densely connected ASPP module,which reduces the amount of calculation and improves the overall segmentation accuracy.Aiming at the problem that the single-level feature fusion does not make full use of low-level features,this thesis fuse two low-level features on the premise of ensuring speed to obtain a better segmentation effect.Aiming at the problem of high-and low-level feature misalignment during feature fusion,this thesis proposes a feature alignment module based on attention mechanism to further improve the segmentation accuracy.Finally,the average crossover on the Cityscapes test set(1024×2048)reached69.4% than mIoU,and the test speed reached 103.3 FPS on the single-card GTX1080 Ti.2.Based on the results of the first part,this thesis studies a lightweight street view image semantic segmentation algorithm with cascading ASPP.For the lack ofcorrelation between the high-level and low-level features of the existing algorithms,and the ASPP module only multi-scale context information aggregation for high-level features,this article uses ASPP instead of ordinary convolution after feature stitching in the decoding stage to perform low-level features.Multi-scale aggregation finally introduces an attention mechanism to further improve the accuracy while meeting the speed requirements.The mIoU on the Cityscapes test set(1024 × 2048)was further increased to 70.0%,and the test speed reached 89.1FPS on the single-card GTX 1080 Ti.3.On the basis of the AM5749 embedded platform,this thesis implements a fast street scene image semantic segmentation algorithm.As for the many types of Cityscapes in the existing data set,this article reorganizes the categories,leaving only the objects that are mainly concerned during driving.Aiming at the problem of limited computing resources of embedded platforms,a lightweight basic network Net14 was designed.The algorithm is sparsely trained by using the caffe-jacinto framework.The sparse rate of the model weights reaches 79%,and then 8-bit quantization is performed to meet the calculation method of the embedded platform.According to the multi-core architecture of AM5749,a pipeline processing is built to further speed up the operation.On the simplified Cityscapes verification set(512×1024),it reached 75.8% of mIoU,and the test speed on AM5749 reached 5.0 FPS.