Improved PPYOLOE for conveyor roller object detection

In mineral processing systems, idlers serve as critical load-bearing components of belt conveyors. However, fault detection in such components is often hindered by real-world challenges, including image blur due to jitter, partial occlusion, background interference, and limited labeled samples. These factors severely impact the accuracy and adaptability of existing detection methods. To address these issues, we propose a decoupled cascade detection paradigm based on a two-stage “object detection–state recognition” framework. This approach prioritizes robust idler localization, providing a solid foundation for subsequent state classification. Specifically, to address the low computational efficiency of PPYOLOE, we introduce a lightweight bottleneck module that significantly reduces parameter count and computational cost while preserving deep semantic expressiveness. To tackle feature incompleteness caused by occlusions, a multi-level gradient path optimization strategy is designed, coupled with a cross-stage feature fusion mechanism to enhance the integration of shallow textures and deep semantics. Furthermore, to mitigate the effects of image blur and annotation scarcity, we construct a dense pseudo-label collaborative distillation framework. A teacher model generates multi-scale pseudo-labels, and a blur-invariance constraint is applied to improve representational robustness under blurred conditions. Experiments conducted on an idler dataset from a mineral processing plant demonstrate that with only 1% labeled data, the proposed pseudo-label distillation framework improves the Average Precision (AP) from 0.5% to 76.9%. At 35% label proportion, AP increases from 85.2% to 85.8%. The enhanced model raises AP from 69.4% to 85.8%, boosts average recall from 74.9% to 88.5%, and increases inference speed from 13.27 to 33.43 FPS. Compared to traditional object detection models, our method significantly improves detection accuracy, generalization, and real-time performance under low-label conditions, offering a superior solution for real-time idler detection in mining belt conveyors.