Research On Joint Deep Methods For Depth Of Field Estimation And Semantic Segmentation

With the rapid development of artificial intelligence research,autonomous driving technology,one of the important indicators of the development level of artificial intelligence,has become a hot research direction in recent years.In the research process in the field of automatic driving,the automatic driving system is usually divided into three modules: perception,planning,and control.This paper mainly discusses the visual problems in the perception module.The vision-based perception module mainly solves the problem of visual scene understanding,which conducts accurate and efficient analysis of the external environment around the vehicle through technical means such as computer vision.Specifically,this paper will focus on two key sub-tasks in the field of autonomous driving scene understanding: depth-of-field estimation and semantic segmentation.In recent years,as deep learning technology has increasingly displayed remarkable role in many fields,scene understanding methods based on convolutional neural networks have gradually shown better performance than traditional methods.However,different from the current widely studied single-task depth model,this paper focuses on the joint depth model for depth-of-field estimation and semantic segmentation based on multi-task learning,mainly including the following research points:Firstly,to deal with the problem that the current multi-task models can’t deeply filter the task sharing features,this paper proposes an end-to-end joint deep learning model for depth-of-field estimation and semantic segmentation.By using the feature sharing module of the Encoder-Decoder structure,the task sharing features are mined,and a novel feature screening module is proposed to extract deep-level features of shared features.Finally,attention mechanisms are introduced.Experimental results show that the proposed model has the characteristics of high accuracy,strong robustness and light weight.Secondly,considering that the current multi-task depth models for scene perception are of great many parameters and hence low efficiency though with high precision,this paper further discusses a joint deep model combining linear multi-step residual module and lightweight technology for depth-of-field estimation and semantic segmentation.Specifically,the linear multi-step residual module is introduced into the Dense multi-task model in the existing literature,and different lightweight modules such as standard convolution,depthwise separable convolution,and the combination of group convolution,depthwise convolution,and channel shuffle operation are exploited.Experimental results show that,with the depthwise separable convolution this paper can reduce the amount of model parameters and computational cost of the original Dense model by 90%without affecting its performance on depth-of-field estimation and semantic segmentation.Finally,the loss of multi-task learning changes dynamically in the training process,so does the relative difficulty between the tasks.The loss of the multi-task model should adaptively adjust the weights between tasks to maintain task balance.To this end,this paper further discusses three kinds of multi-task loss function weight adjustment schemes,and makes comparative experiments using the multi-task models proposed in the paper.Experimental results show that the multi-task loss function weight adjustment scheme based on uncertainty has the best effect under different multi-task models and different scenarios.