Road Detection Method Based on Soft Cross-layer Connection and Self Attention
Self attention is one of the important methods to improve the accuracy of road detection.Compared with attention mechanism,self attention reduces the dependence on external information and makes it have stronger data capture ability.When cal-culating the response between two locations,Non-local network will consider the weight of all dimensional feature graphs,but the amount of calculation will increase.SENet obtains the weight matrix with global receptive field through extrusion and excitation,but the fitting ability of SENet to road edge is limited.A new self attention module is designed to compress the information in different di-mensions by convolution and matrix multiplication without increasing too much computation,so as to obtain more global informa-tion.Flexible cross layer connection is introduced between encoder and decoder to improve network performance.In terms of loss function,BCD(Bray Curtis distance)and hinge loss are introduced based on the common cross entropy loss function.By giving higher punishment to the wrong points near the road edge,the network has the ability to better fit the road edge.Experiments show that the detection accuracy of this method is higher than that of Non-local net and SENet.
国家科技期刊平台
NSTL
万方数据
