MSBG-DeepLabV3+ Conveyor Belt Idler Roller Segmentation Algorithm

Abstract

Idler rollers play a crucial supporting and guiding role in the operation of conveyor belts, and their state changes significantly affect conveyor belt misalignment. Studies have found a significant difference in the exposed area of the left and right idler rollers under misalignment and non-misalignment conditions. Based on this characteristic, this paper focuses on idler rollers and proposes a semantic segmentation algorithm based on MSBG-DeepLabV3+ idler rollers to achieve high-precision segmentation of the idler roller region, providing technical support for subsequent research on conveyor belt misalignment. Building upon the original DeepLabV3+ framework, firstly, MobileNetV2 is used to replace the Xception backbone network, achieving model lightweighting. Secondly, a strip perception DenseASPP is introduced in the encoding stage to enhance the multi-scale contextual modeling capability of the slender directional structure of the idler rollers. Finally, in the decoding stage, a boundary enhancement module (BEM) and a global directional attention mechanism (GDAM) are combined to further improve the segmentation boundary accuracy and the perception capability of key regions and directional features. Experimental results show that the proposed MSBG-DeepLabV3+ model achieves an mIoU of 97.16% and an mPA of 98.62% on a self-made dataset. While ensuring high detection efficiency, this method achieves high-precision and high-efficiency segmentation of idler roller targets in complex industrial environments, verifying the effectiveness and practicality of the proposed algorithm.