Adaptive edge detection based on multi scale residual networks

DOI:	10.11809/bqzbgcxb2023.07.003
Keywords:	edge detection; adaptive threshold; multi scale feature fusion; ResNet18; UAV
Abstract:	In order to realize obstacle detection in the process of UAV obstacle avoidance, this paper proposes an adaptive edge detection algorithm based on multi scale residual networks to deal with slow detection speeds, poor noise resistance performance and easy loss of edge information of traditional edge detection algorithms. Based on deep separable convolution, the traditional convolutional structure in the ResNet18 residual network is replaced to reduce the number of network parameters and improve the efficiency of feature extraction. On the basis of a comprehensive consideration of fixed thresholds and local edge point means, adaptive thresholds are introduced to improve the eight way difference operators so as to improve the ability of the algorithm to suppress noise. Furthermore, a multi scale feature fusion structure is designed to fuse low level feature information on the high level feature map with stronger semantic information and refine the small scale target edge features. Compared with the unimproved algorithm, the accuracy rate increases by 4.23%, the number of parameters reduces by 5.01×106, and the processing speed increases by 4 fps.
Issue:	Vol. 44 No. 7 (2023)
Published:	2023-07-28
PDF HTML