针对使用深度卷积神经网络识别煤中异物的推理速度受其高计算成本的阻碍的问题,提出了一种基于深度学习的煤中异物识别模型轻量化策略,减小了模型大小和运行时的占用内存,并能在不影响准确性的情况下降低计算操作数量。该方法将L1正则化强加给批量归一化(BN)层中的比例因子,使得能够自动识别不重要的通道或神经元。不重要的通道会被自动识别并在之后进行修剪,从而产生具有相当精度的精简模型。该轻量化策略可提高煤矿工业生产效率和安全性,并为其他领域的深度学习模型提供参考。

In response to the problem that the inference speed of using deep convolutional neural networks to identify impurities in coal is hindered by their high computational cost, we present a lightweight strategy for coal impurity identification based on deep learning, aiming to simultaneously reduce model size, decrease memory usage during runtime, and reduce the number of computational operations without compromising accuracy. This approach enforces L1 regularization on the scale factors within Batch Normalization (BN) layers, allowing automatic identification of unimportant channels or neurons. Unimportant channels are automatically identified and pruned, resulting in a compact model with comparable accuracy, significantly reducing model size and computational overhead. This means that our approach can enhance the efficiency and safety of the coal mining industry. Furthermore, the lightweight strategy holds reference value for deep learning models in other domains.