Abstract: The coal mining face of the mine has a long distance from the ground, and needs to be transported over long distances by coal transport belts, and in the process of transport, there are problems such as large gangue, anchors and other foreign objects damaging the belts and blocking the coal drop openings, which are prone to cause safety problems, so the classification of foreign objects in the transport of coal transport belts is of great significance. In order to overcome the problems of weak light intensity, low recognition accuracy and large number of model parameters in the underground environment, an XDNet-CFF lightweight network combined with low-light image enhancement is proposed. Firstly, a pre-trained self-calibrated light image enhancement model is used to perform low-light image enhancement on underground coal transport belt images to effectively improve the image quality; secondly, a deep network based on Xcpetion-DenseNet121 and cross-layer feature fusion is designed to improve the feature extraction capability while combining the underlying detail features with the upper layer semantic features to reduce the loss of information and enrich the feature representation; then, the coal belt foreign object recognition is completed by fully connected layer and softmax; finally, in order to achieve mobile deployment and recognition warning, the model is compressed by applying pruning method, which significantly reduces the number of model parameters and overheads. The results show that the proposed model achieves accuracy, precision, recall and F1 score of 0.9467, 0.9512, 0.9416, 0.9464 on the coal belt foreign object dataset, which are better than the mainstream model, and at the same time, the number of references is only 8.98 M, which meets the demand of actual production deployment.