关键词: 煤岩识别, 语义分割, 少样本学习, U-net, 深度学习, 机器视觉技术

Abstract: In recent years,image semantic segmentation methods have been widely applied in coal rock interface recognition research. However,currently labeled coal rock image data samples are difficult to obtain,and there is a lack of public datasets. Moreover,existing semantic segmentation models usually rely on large sample datasets for training. In response to the above issues,this article proposes a small sample coal rock interface image segmentation model based on an improved U-net model. Firstly,the VGG16,which has stronger feature extraction capability and simpler structure,is used as the backbone network,which can enhance image feature extraction efficiency and achieve faster training speed. Secondly,during the training of the improved network model,the transfer learning method is adopted to improve the model accuracy and avoid overfitting,making the model more suitable for training with small sample datasets. Additionally,the attention mechanism module is introduced in the skip connections and upsampling section of the U-net network,which helps the model capture key features,enhances the model′s feature extraction capability,and improves the accuracy of coal-rock interface image segmentation. The performance of the improved network model in this study is verified using a homemade coal-rock interface dataset. By comparing this model with the classical U-net,DeepLabv3+,PSPnet,and HRNet network models,experiment results show that under the same training conditions using a small sample dataset constructed from 125 coal-rock interface images,the improved model in this study has a significant improvement in segmentation accuracy and detection speed compared to the original U-net model. The model′s accuracy has improved by 1. 84%,the mean intersection over union has increased by 5. 34%,the average pixel accuracy of the class has increased by 0. 48%,and the detection speed has increased by 5. 3%. At the same time,compared with other network models,the proposed improved model has a significant advantage in the semantic segmentation of small-sample coalrock interface images,which shows the effectiveness of the proposed improved approach.

Key words: coal-rock recognition,semantic segmentation,few-shot learning,U-net,deep learning,machine vision technique