矿井中开采出来的煤炭要经过运煤皮带的长距离运输才能到达地面。大量有关矿井中煤炭安全高效运输的研究显示，皮带在煤炭输送过程中存在大块矸石、锚杆等异物划伤、撕裂皮带和堵塞落煤口等安全隐患，预警、分选及联动控制不及时会严重影响煤炭的运输效率。为克服当前对皮带异物分类识别时存在的网络参数量大、实时性差、识别精度低等问题，提出了一种融合残差信息的轻量级网络。该网络采用残差块作为基本特征提取单元，在残差块中去除卷积块之间的激活函数。采用交叉学习机制和特征拼接的方法来融合不同尺度的特征信息，增强了特征的表现力。精简信息融合网络的结构并增加信息融合网络的数量，提高了模型的扩展性。在模型进行前向传播时，对损失函数进行阈值处理，改善了测试集损失函数升高的问题，提高了模型的泛化性。提出的轻量级分类网络模型在Cifar10数据集、Cifar100数据集和矿用数据集的分类准确率分别为94.1%，73.9%和85.1%。在矿用数据集上与ShufflenetV2，MobileNetV2，ResNet50，ResNeXt50，W-ResNet50，ResNet110等算法相比，本文提出的模型的识别准确率分别提升了4.2%，4.3%，0.7%，0.5%，0.3%和0.8%；此外，与分类准确率同本文网络相近的ResNet50，ResNeXt50，W-ResNet50和ResNet110算法相比，FPS分别提高了28，26，34和46。结果表明本文算法在提高对运煤皮带的异物进行分类识别精度的同时，其计算速度也得了提升，显著提高了煤炭的运输效率，促进了计算机视觉与煤矿安全生产的深度融合。

The coal in the mine must be transported long distance by coal belts before reaching the ground. Plenty ofstudies on the safe and efficient transportation of coal in mines reveal that the belts often suffer some hazards caused byforeign objects such as large gangue, bolts and other foreign bodies scratching, tearing the belt, and blockingthe coal discharge point in the process of coal transportation. If the early warning, sorting and linkage control are not timely, it will seriously affect the coal transportation efficiency. To overcome the problems of large amountof network parameters, poor real⁃time performance, and low recognition accuracy in the current classification and rec⁃ognition of belt foreign objects, a lightweight network that integrates residual information is proposed. Firstly, the re⁃sidual block is adopted as the basic feature extraction unit of the network, and the activation function is removed be⁃tween the convolution blocks in the residual block. Then, the cross⁃learning mechanism and feature splicing methodare used to fuse the feature information of different scales, resulting in the enhanced expressiveness of the features.Furthermore, the structure of the information fusion network is simplified and the number of information fusion net⁃works is increased, which improves the scalability of the model. Moreover, the loss function is thresholded during theforward propagation of the model, which can solve the problem of elevated test set loss function and improve the gener⁃alization of the model. By conducting the experiments on the Cifar10, Cifar100 and the mining dataset, the recognitionaccuracy of the proposed network model can reach as high as 94.1%, 73.9% and 85.1%, respectively. Comparedwith the ShufflenetV2, MobileNetV2, ResNet50, ResNeXt50, W⁃ResNet50 and ResNet110 algorithms on the min⁃ing dataset, the accuracy rates proposed are 4.2%, 4.3%, 0.7%, 0.5%, 0.3% and 0.8% higher than those respec⁃tively. In addition, compared with ResNet50, ResNeXt50, W⁃ResNet50 and ResNet110, whose classification accura⁃cies are similar to the network proposed, the FPS can be increased by 28, 26, 34 and 46, respectively. The re⁃sults demonstrate that while improving the classification and identification accuracy of foreign objects, the calculationspeed of the proposed algorithm in this paper has also been accelerated, which can significantly improve the transpor⁃tation efficiency of coal and promote the deep integration of computer vision and coal mine safe production.