ABSTRACT

A method of identifying and positioning coal and gangue on mobile devices is provided to address the difficulties of high complexity and difficult deployment of existing machine vision algorithms. Combining CSPDarknet53 and the GhostNet embedded with efficient channel domain attention to build a lightweight feature extraction network and using the Meta-ACON activation function to adapt the network’s nonlinearity at each layer. The detection head’s expression capability is enhanced using a decoupled head. The actual position coordinate relationship equation of coal and gangue is deduced by borrowing the center point object box positioning principle. Build an experimental setup, collect and construct coal and gangue datasets, and use Mixup data augmentation to improve the network’s ability to position coal and gangue for stacking cases. The models are trained, and tested for different lighting conditions, sizes, and stacking cases, then deployed on the mobile device. The results show that the model has the highest detection accuracy and can correctly identify and position both coal and gangue under complex conditions with a high confidence value. The FLOPs is 76.83% lower compared to the original network YOLOv5s. The mAP reaches 0.996 and the FPS is 40.11. The inference time on the mobile device is reduced to 228 ms, which basically meets the speed requirement of coal gangue identification and positioning.

摘要

为解决现有机器视觉算法复杂度高、部署困难的问题，提出了一种在移动设备上识别定位煤矸石的方法。结合 CSPDarknet53 和嵌入高效通道域注意力的 GhostNet 构建轻量级特征提取网络，并使用 Meta-ACON 激活函数来适应网络在每一层的非线性。使用解耦头增强了检测头的表达能力。借用中心点物盒定位原理，推导出煤与煤矸石的实际位置坐标关系方程。建立实验装置，收集和构建煤和煤矸石数据集，并使用 Mixup 数据增强来提高网络定位煤和煤矸石以进行堆放的能力。这些模型针对不同的光照条件、尺寸和堆叠情况进行训练和测试，然后部署到移动设备上。结果表明，该模型具有最高的检测精度，能够在复杂条件下对煤和煤矸石进行正确识别和定位，具有较高的置信度值。与原始网络 YOLOv5s 相比，FLOPs 降低了 76.83%。mAP达到0.996，FPS为40.11。移动端推理时间降低至228ms，基本满足煤矸石识别定位的速度要求。结果表明，该模型具有最高的检测精度，能够在复杂条件下对煤和煤矸石进行正确识别和定位，具有较高的置信度值。与原始网络 YOLOv5s 相比，FLOPs 降低了 76.83%。mAP达到0.996，FPS为40.11。移动端推理时间降低至228ms，基本满足煤矸石识别定位的速度要求。结果表明，该模型具有最高的检测精度，能够在复杂条件下对煤和煤矸石进行正确识别和定位，具有较高的置信度值。与原始网络 YOLOv5s 相比，FLOPs 降低了 76.83%。mAP达到0.996，FPS为40.11。移动端推理时间降低至228ms，基本满足煤矸石识别定位的速度要求。