煤矸井下智能分选作为智慧矿山建设的重要组成，可有效提升矿井资源绿色利用。现阶段可见光图像识别技术针对井下昏暗潮湿环境中煤矸混合体的辨识还有待完善。基于热红外成像技术和改进YOLOv5算法模型，提出了一种暗湿工况下煤矸混合态势热敏图像辨识方法。将YOLOv5模型的Neck部分改用加权双向特征金字塔(BiFPN)结构，通过多层次特征融合提高煤矸的辨识效率，采用CIOU函数作为损失函数，提升煤矸检测精准率；构建了煤矸混合体热敏采集实验平台，模拟了井下密闭空间低照度、高湿度、高风速环境，通过CLAHE与LAPLACE算子对红外摄像机所采集的热敏图像进行对比度增强和边缘强化预处理，以不同数据集、不同改进模块、不同算法模型等多个角度系统分析了煤矸混合体态势热敏图像辨识结果，探究了湿度变化对暗湿工况下煤矸识别准确率的影响规律。研究结果表明：预处理后的图像平均精准率较原始图像提升了1.7%，F-Measure提升了6.9%；改进后的YOLOv5模型平均精度均值和F-Measure达到了80.2%与84.6%，高于经典模型的74.6%与79.7%，可有效提升煤矸热敏图像检测精度；环境相对湿度与识别准确率呈现先正相关待湿度达到一定阈值后负相关的变化规律。提出了热敏图像可准确识别昏暗潮湿密闭环境中的煤矸混合体，为实现井下暗湿工况煤矸混合态势精准辨识提供了科学依据。

As an important component of intelligent mine construction, the underground intelligent separation of coal and gangue can effectively improve the green utilization of mine resources. At present, the visible light image recognition technology for the identification of coal-gangue mixture in dark and humid underground environment remains to be improved. Based on the thermal infrared imaging technology and improved YOLOv5 algorithm model, this paper proposed a thermal image identification method for the coal-gangue mixed situation under dark and wet conditions. The neck part of the YOLOv5 model was changed to Bidirectional Feature Pyramid Network (BiFPN) structure, and the identification efficiency of coal-gangue was improved through multi-level feature fusion. The CIOU function was used as the loss function to improve the accuracy of coal-gangue detection. An experimental platform for the thermal image acquisition of coal-gangue mixture was constructed to simulate the low illumination and high humidity environment in underground confined space. The contrast enhancement and edge enhancement preprocessing of the thermal image collected by the infrared camera were performed by CLAHE and LAPLACE operators. The results of thermal image identification of coal gangue mixture were systematically analyzed from different data sets, different improved modules and different algorithm models, and the influence of humidity change on the accuracy of coal-gangue identification under dark and wet conditions was explored. The results show that the average accuracy of the preprocessed image is 1.7% higher than that of the original image, and the F-Measure value is increased by 6.9%. The average accuracy mean and F-Measure of the improved YOLOv5 model reaches 80.2% and 84.6%, which are higher than 74.6% and 79.7% of the classical model, which can effectively improve the detection accuracy of coal gangue thermal image. The relative humidity of the environment is positively correlated with the recognition accuracy, and negatively correlated after the humidity reaches a certain threshold. It is proposed that the thermal image can accurately identify the coal-gangue mixture in the dark and humid closed environment, which provides a scientific basis for the accurate identification of the coal-gangue mixed situation under the dark and wet conditions.