为解决人工统计钻孔冲煤量不准确以及效率低等问题，提出一种YOLOv8n、ResNet34和PP-OCRv4算法相结合的智能识别方法。

该方法首先使用YOLOv8n算法完成一级检测，同时并行级联ResNet34算法与PP-OCRv4算法进行二级处理，并结合基于追踪帧数的分类状态判别方法，建立了冲煤量自动计算的算法框架。其次，在YOLOv8n的C2f模块中引入可变形卷积DCNv2模块，以削弱点状强光照对特征采集的影响，并将其默认的检测头替换为Dynamic Head检测头模块，以强化算法在尺度，空间和通道维度的特征提取能力，以及将CIoU损失函数替换为SIoU损失函数，以加速预测框与真实框的匹配，并利用自建的数据集对改进后的YOLOv8n算法进行验证。

结果表明：(1)与原算法相比，平均类别检测精度提高了7.6%，召回率提高了3.5%，精确率提高了6.4%，验证了改进策略对提升模型性能的有效性和稳定性。(2)对4个不同的瓦斯抽采水力冲孔钻场的实时视频进行测试，识别准确率分别为100.0%、93.3%、95.7%和93.1%，平均达到95.5%，满足了水力冲孔钻孔冲煤量自动识别的精度要求。(3)采用追踪帧数确定ResNet34分类状态的方法，解决了分类状态单次识别结果不可靠的问题。研究成果为YOLO系列算法与其他深度学习技术的融合和广泛应用提供了技术与实践基础，对促进瓦斯抽采钻场等煤矿井下复杂场景的智能化进步具有参考价值。

Given the inaccurate and low-efficiency manual statistics of the quantity of coals flushed out during borehole hydraulic flushing, this study proposed an intelligent identification method that combines YOLOv8n, ResNet34, and PP-OCRv4 algorithms.

First, the first-level detection was completed using the YOLOv8n algorithm, which, combined with the ResNet34 and PP-OCRv4 algorithms via parallel cascade, was then employed to conduct second-level processing. Through the above processing and in combination with the method for classification status discrimination based on tracking frame numbers, the framework of an algorithm for automatically calculating the quantity of coals flushed out was established. Subsequently, the deformable convolution DCNv2 module was introduced into the C2f module of YOLOv8n to reduce the impacts of point-like strong illumination on feature extraction. Moreover, the default detection head of YOLOv8n was replaced with the Dynamic Head module to strengthen the feature extraction in scale, space, and channel dimensions. The CIoU loss function was replaced with the SIoU loss function to accelerate the matching between prediction and ground truth boxes. Finally, the improved YOLOv8n algorithm was validated using a self-built dataset.

The results indicate that, compared to the original algorithm, the improved YOLOv8n algorithm increased the mean average classification accuracy, recall, and precision by 7.6%, 3.5%, and 6.4%, respectively. This verifies the effectiveness and stability of the improvement strategy for enhancing the model performance. The improved YOLOv8n algorithm was applied to the real-time videos from four drilling sites of borehole hydraulic flushing for gas drainage, yielding respective identification accuracies of 100.0%, 93.3%, 95.7%, and 93.1%, with an average of 95.5%, meeting the accuracy requirements for the automatic identification of coal quantity flushed out during borehole hydraulic flushing. The method for determining the ResNet34 classification status based on tracking frame numbers resolved the problem of unreliable single identification of the classification status. The results of this study provide a technical and practical foundation for the integration of the YOLO series of algorithms with other deep learning techniques and its wide applications. Besides, these results serve as a valuable reference for achieving intelligent advances in complex underground coal mine scenarios such as drilling sites for gas drainage.