Weakly Supervised Road Garbage Quantification System Based on WCCA

Abstract

Mechanized road sweeping vehicles are an indispensable part of urban infrastructure. However, while improving work efficiency, the high energy consumption and noise pollution of their high-power dust collectors have become new problems that need to be solved. To achieve adoptive control of dust collection power, we propose a vision-based road sweeper intelligent power control system. The system use the proposed Weight Criss-Cross Attention (WCCA) module, embed into the Mobile V2 light-weight image-level classification network, to achieve weakly supervised pixel-by-pixel segmentation of road garbage area, and finally get the road garbage quantification result. With the proposed prior loss function based on the distribution of road garbage image data, WCCA can guide the convergence direction of the model to the correct target area. Finally, two state-of -the-art comparison algorithms are used to prove the superiority of the proposed algorithm.