ABSTRACT

With the development of machine vision and deep learning, the intelligent visual-based coal-gangue separation technology has gradually attracted the attention of enterprises and researchers. Coal-gangue detection models based on deep learning rely on a large number of CG¹1. CG: coal and gangue images for training, which is time-consuming and laborious to acquire. In this paper, based on small samples, image augmentation, and generative adversarial network were used to expand dataset of CG images to improve the performance of coal-gangue detectors. The dataset was expanded by pixel transform firstly. Next, four improved DCGAN structures were proposed to generate more diversiform CG images. Refer to the results of model training and generated image evaluation, DCGAN32 with three blocks had the best ability to generate more authentic image with the lowest FID value, scoring 93.8708 and 104.3394 on coal and gangue, respectively. As the training data expands, the performance of the object detection model was improved by up to 9.4% and 15.1%, respectively, on mAP and IoU. The proposed data expansion scheme can effectively improve the training performance of coal-gangue detectors based on small samples.

摘要

随着机器视觉和深度学习的发展，基于智能视觉的煤矸石分离技术逐渐引起了企业和研究人员的关注。基于深度学习的煤矸石检测模型依赖大量的CG ¹1.CG：煤和矸石用于训练的图像，获取这些图像既费时又费力。在本文中，基于小样本，图像增强和生成对抗网络用于扩展 CG 图像数据集，以提高煤矸石探测器的性能。首先通过像素变换扩展数据集。接下来，提出了四种改进的 DCGAN 结构来生成更多样化的 CG 图像。参考模型训练和生成图像评估的结果，具有三个块的 DCGAN32 以最低的 FID 值生成更真实的图像的能力最好，煤和矸石的得分分别为 93.8708 和 104.3394。随着训练数据的扩大，目标检测模型在 mAP 和 IoU 上的性能分别提高了 9.4% 和 15.1%。