An EfficientNetv2-based method for coal conveyor belt foreign object detection

Abstract

Listen

Translate article

The detection and recognition of foreign objects on coal conveyor belts play a crucial role in coal production. This article proposes a foreign object detection method for coal conveyor belts based on EfficientNetv2. Since MBConv and Fused-MBConv structures in EfficientNetv2 employ information compression and fusion strategies, which may lead to the loss of important information and affect the integrity of feature extraction, a hard shuffle attention (Hard-SA) mechanism is utilized to enhance the focus on important features and improve the representation ability of coal conveyor belts image features. To address the potential gradient disappearance issue during the backpropagation process of the network, an elastic exponential linear unit (EELU) activation function is introduced. Additionally, since the cross-entropy loss function may not be flexible enough to handle complex data distributions and may fail to fit the non-linear relationships between data well, a Polyloss function is adopted. Polyloss can better adapt to the complex data distribution and task requirements of coal mine images. The experimental results show that the proposed method achieves an accuracy of 93.02%, which is 2.39% higher than that of EfficientNetv2. It also outperforms some other state-of-the-art (SOTA) models and can effectively complete the detection of foreign objects on coal conveyor belts.