In the process of coal mining, the coal quality is greatly reduced due to the mixing of gangue, and the stacking and burning of coal and gangue can cause serious environmental pollution. The fast identification of coal and gangue is an important technology for gangue sorting. Considering tiny YOLO v3 has the advantages of fast running speed, simple network and good effectiveness, an improved tiny-YOLO-v3-based fast identification model is proposed in this paper, which including the spatial pyramid pooling (SPP) net, the squeeze-and-excitation (SE) module and the dilated convolution. Firstly, the SPP net preprocesses the input images into acceptable size via a single convolution layer. Then, the SE module, which can strengthen the attention among channels in RGB image, is adopted to capture the key information and enhance the sensitivity of network accurately. Finally, the dilated convolution, which can enlarge the receptive filed without increasing parameters, is used to further optimize and realize the fast identification of coal and gangue. The experiments show that, compared with the tiny YOLO v3 the average intersection over union (Avg IOU) of proposed model is as high as 0.39%, the required time for each image and the loss reduced by 7.41 and 53.01%, respectively. The mean Average Precision (mAP@0.5) of the proposed network is 3.12% higher than Faster RCNN, reaching 0.964. It is an efficient fast identification model for gangue-sorting.

在煤炭开采过程中，由于矸石的混入，煤质大大降低，煤与矸石的堆放燃烧会造成严重的环境污染。煤矸石的快速识别是矸石分选的一项重要技术。考虑到tiny YOLO v3具有运行速度快、网络简单、效果好等优点，本文提出一种改进的基于tiny-YOLO-v3的快速识别模型，包括空间金字塔池化（SPP）网络、squeeze-和激励（SE）模块和空洞卷积。首先，SPP 网络通过单个卷积层将输入图像预处理为可接受的大小。然后，SE模块，可以加强RGB图像通道间的注意力，用于准确捕捉关键信息，增强网络的灵敏度。最后利用扩张卷积在不增加参数的情况下扩大感受野，进一步优化实现煤矸石的快速识别。实验表明，与微型 YOLO v3 相比，所提出模型的平均交叉并集（Avg IOU）高达 0.39%，每张图像所需的时间和损失分别减少了 7.41% 和 53.01%。提出的网络的平均平均精度（mAP@0.5）比 Faster RCNN 高 3.12%，达到 0.964。它是一种高效的矸石分选快速识别模型。用于进一步优化和实现煤矸石的快速识别。实验表明，与微型 YOLO v3 相比，所提出模型的平均交叉并集（Avg IOU）高达 0.39%，每张图像所需的时间和损失分别减少了 7.41% 和 53.01%。提出的网络的平均平均精度（mAP@0.5）比 Faster RCNN 高 3.12%，达到 0.964。它是一种高效的矸石分选快速识别模型。用于进一步优化和实现煤矸石的快速识别。实验表明，与微型 YOLO v3 相比，所提出模型的平均交叉并集（Avg IOU）高达 0.39%，每张图像所需的时间和损失分别减少了 7.41% 和 53.01%。提出的网络的平均平均精度（mAP@0.5）比 Faster RCNN 高 3.12%，达到 0.964。它是一种高效的矸石分选快速识别模型。