关键词: 深度学习, 单目图像深度估计, 多尺度注意力导向网络, 多尺度特征, 通道注意力融合

Abstract:

Aiming at the problems of low spatial resolution and unclear edges in the existing depth estimation algorithms of monocular images based on deep learning，a depth estimation algorithm of monocular images based on multi-scale attention-oriented network was put forward． Firstly，an end-to-end encoder-decoder model was designed，and the encoder extracts features at multiple scales． To ensure better depth continuity，the decoder gradually optimize details and scene structure of extracted multi-scale features by combining residual learning with channel attention fusion． Considering the loss of depth details caused by multiple down-sampling，a boundary enhancement module was designed． By introducing spatial attention，the inter-class contrast of different objects was improved to enhance the boundary details of the image． Finally，the optimization module fuses multi-scale features from the decoder and the boundary enhancement module to generate a depth image． Experimental results show that，compared with the current mainstream algorithms，the depth image generated by the algorithm has improved quality，showing more detailed object contour information and good performance in both objective indicators and subjective effects．

Key words: deep learning, monocular image depth estimation, multi-scale attention-oriented network, multi-scale feature, channel attention fusion