An adaptive focused target feature fusion network for detection of foreign bodies in coal flow

Ye, Tao

DOI: https://doi.org/10.1007/s13042-023-01798-6

2023-03-12

International Journal of Machine Learning and Cybernetics

Abstract:In the process of conveying raw coal to the surface on conveyor belts, the raw coal is generally blended with foreign bodies, such as large pieces of gangue and damaged bolts, which can affect the quality of mined coal, damage the transportation equipment and even jam the coal conduit, seriously reducing the coal conveying efficiency. To handle to the existing problems of underground complex environment, low detection accuracy and poor real-time performance in coal flow foreign bodies detection, we propose an adaptive focused target feature fusion network (AFFNet) based on YOLOX. The multi-transformer parallel (MTRP) module is used to expand the receptive field and fuse the features under different receptive fields with the transformer encoder to enhance the feature extraction ability. The cross stage partial transformer (CSPTR) with transformer encoder module is designed to capture the global context information of the feature maps in the network, and improve the location prediction in the detection. In the feature fusion channel of different scales, the learnable weight parameters are added to learn the spatial weight of feature map fusion adaptively, and the feature expression ability of different scales is optimized. SCYLLA-IoU (SIoU) loss and varifocal loss are used to obtain more accurate bounding boxes and deal with the sample category imbalance problem, respectively. The experimental results show that AFFNet can achieve a detection speed of 48 frame per second (FPS) and a mean average precision (mAP50) of 95.6%, 6.7% higher than YOLOX-s on the dataset of foreign body in coal flow. It can balance both the detection speed and detection accuracy, and can be used to improve the efficiency of detecting foreign bodies in the coal flow.

computer science, artificial intelligence