A window-based multi-scale attention model for slope collapse detection

Abstract

Slope collapse detection is a critical foundation for defending against water flow impact and maintaining shoreline stability. Under realistic detection conditions such as difficult collection of actual defect samples and complex target environment, addressing the problem of existing deep learning techniques with low recognition accuracy, a window-based multi-scale attention model for slope collapse detection called Faster-Swin is proposed. The traditional convolutional neural network (CNN) architecture has been enhanced by incorporating skip connection structures such as Elan block and SPPC block, which alleviate the problem of gradient vanishing that arises from increasing depth in deep neural networks. Then the window-based multi-scale attention mechanism is added to the backbone network to increase the overall modeling ability of the model and reduce the computational cost. Experimental results show that Faster-Swin can adapt to complex backgrounds in low-quality datasets, and the final mean average precision (map) value can reach 97.16%.