A Multi-Scale Perception Feature Fusion Method for Self-Supervised Monocular Depth Estimation

Self-supervised monocular depth estimation is a strategy that can be trained without requiring ground truth. For precise depth prediction, current techniques typically fall short of fully using the encoder-decoder to extract possible characteristics from the constrained sensory field. A hybrid module DGFE based on a convolutional neural network and attention mechanism is proposed for this purpose in response to the inefficiency of self-supervised monocular depth estimation in extracting features. This module is integrated into the encoder of the network GOHA designed in this paper to enhance the network's ability to extract and represent features. Furthermore, an enhanced pooling operation is presented in addition to the traditional pooling layer operation. The self-supervised monocular depth estimation approach described in this research has improved several performance metrics, according to experimental results.