煤矸石分选是保障煤炭清洁、高效利用的一个重要环节。为顺应煤矸分选技术自动化、智能化的发展趋势，本团队设计了一种煤矸分拣机器人，该机器人应用图像目标检测技术识别、定位煤矸石，使用气动机械手爪抓取矸石完成排矸任务。煤矸石图像目标检测是煤矸分拣机器人进行自动分选的关键技术之一，传统检测方法侧重于对煤矸石图像识别的研究，存在图像处理过程复杂、识别算法应用场景受限等问题。针对上述问题，本文在分析总结前人研究工作的基础上，提出采用深度学习技术进行煤矸石目标检测的方法，解决煤矸石的识别与定位问题，为机器人分选排矸提供依据。主要研究内容如下：

针对目前尚不存在一个公开可用的煤矸石数据集的问题，依据PASCAL VOC数据集中的图像样本特征与格式，收集不同场景下的煤矸石图像样本，采用筛选、样本扩充等方式预处理图像，在此基础上标注图像并构建煤矸石数据集。

根据煤矸石检测模型的选取要求，提出基于Faster R-CNN算法的煤矸石图像检测方法，并给出煤矸石检测的整体流程。为了减少煤矸石检测模型的训练时间和资源，采用迁移学习方法预训练模型，在此基础上应用煤矸石数据集训练得到基于Faster R-CNN算法的煤矸石检测模型。与SSD煤矸石检测模型进行性能对比实验，验证本文煤矸石检测方法的可行性与可靠性。

通过研究Faster R-CNN网络先生成目标推荐区域、再进行分类定位的检测原理，提出使用ResNet-50代替原算法中的VGG作为特征提取网络，并采用Soft-NMS策略代替NMS策略筛选边框，以提升煤矸石检测性能。设置不同锚框的种类和数量，验证适当减少锚框的种类和数量能够在提高模型检测速度的同时保证较好的检测精度。

在煤矸分拣机器人系统中，搭建视觉识别模块，该模块可实现检测数据的采集、图像预处理、煤矸石在线检测和检测效果展示等功能。利用训练好的煤矸石检测模型对不同环境下的254个煤矸石样本进行在线检测实验与分析，针对误检、漏检情况提出改善措施，并且以检测框中心对目标矸石进行定位验证。实验结果表明，本文提出的煤矸石检测模型在煤矸分拣机器人系统的应用中取得了良好的检测效果，符合研究预期。

The sorting of coal gangue is an important segment to ensure clean and efficient utilization of coal. In order to conform with the development trend of automation, intelligence of coal gangue sorting technology, a coal-gangue sorting robot has been designed. It uses image object detection technology to identity and locate coal gangue, and uses mechanical paws to grasp gangue to separate coal gangue. Coal gangue image detection is one of the key techniques for automatic separation of the coal-gangue sorting robot. Traditional detection methods mainly focus on the research of coal gangue image recognition, which has the problems of complex image processing and limited application scenarios of recognition algorithms. In view of those problems and based on previous researches, this paper proposes a method of using deep learning to detect coal gangue to solve the problem of coal gangue recognition and positioning, and provide basis for the coal-gangue sorting robot to separate coal and gangue. The main research contents are as follows:

At present, there is a lack of publicly available coal gangue data set. According to image features and format of the PASCAL VOC, coal gangue images in different scenarios are collected, and the images are preprocessed by means of screening, sample augment, etc. On this basis, the processed images are labeled and then a coal gangue data set is created.

According to the selection requirements of coal gangue detection model, this paper proposes a coal gangue detection method based on Faster R-CNN and gives the overall process of coal gangue detection. In order to reduce training time and resources of the coal gangue detection model, the transfer learning is used to pre-train the model, then the coal gangue detection model based on Faster R-CNN is trained with the coal gangue data set. Compared with the SSD gangue detection model by experiments, this paper verifies the feasibility and reliability of the coal gangue detection method.

By studying the detection principle of Faster R-CNN to first generate region proposal and then classify and positioning, it is proposed to use ResNet-50 replace VGG of the Faster R-CNN algorithm as the feature extraction network, and use Soft-NMS instead of NMS to filter the border, to improve the detection performance. Setting different types and numbers of anchor, and it is verified through experiments that appropriate reduction of types and numbers of anchor can improve the detection speed while ensuring better detection accuracy.

In the coal-gangue sorting robot system, a visual recognition module is built, which can realize functions of collection of detection data, image preprocessing, online detection of coal gangue, detection effect display, etc. The trained coal gangue detection model is used to online test 254 coal gangue samples in different environments, and improvement measures are proposed for the false detection and missed detection, and the target gangue is positioned at the center of the rectangular detection frame. The experimental results show that the model proposed in this paper has achieved good detection effect in the application of the coal gangue sorting robot system, which meets the research expectations.