基于深度学习的普通金属矿石快速分拣系统的研究

RESEARCH ON RAPID SORTING SYSTEM OF COMMON METAL ORE BASED ON DEEP LEARNING

刚开采的金属矿石存在大量泥块混合物,需要进行重复的洗矿操作,耗费大量水资源和机器资源,对于粘结性强的泥块还需要后期人工分拣。传统方法一般是通过改造洗矿工艺来减少含泥量,但是改造成本高,不能有效应对较大的泥块,为此提出一个基于深度学习的矿石粗分拣的系统。通过普通RGB摄像头实时采集矿泥混合物图像,并无线传输到服务器端;引入SSD512目标检测框架作为系统核心算法,对矿石和泥块进行识别,并返回目标分类置信度和位置信息的回归;由识别结果控制分拣装置分拣出泥块。实验验证结果发现识别的平均精确率达到91%,识别的速度达到了0.05 s每幅图片。

There is a large amount of mud mixture in the newly mined metal ore, which needs repeated ore washing operation and consumes a lot of water resources and machine resources. In the process of ore washing, the clay with strong cohesiveness needs manual sorting in the later stage. The traditional method is to reduce the mud content by reforming the ore washing technology, but the cost is high and it can’t deal with the larger mud effectively. Therefore, a rough ore sorting system based on deep learning is proposed. The image of the mixture of the ore and mud was collected by the ordinary RGB camera in real time and transmitted to the server by wireless. The SSD512 target detection framework was introduced as the core algorithm of the system to identify the ore and return the confidence degree of the target classification and the regression of the location information. The mud was sorted out by the sorting device controlled by the recognition results. The experimental results show that the average accuracy of recognition is 91%, and the recognition speed is 0.05 s per image.