Image recognition method of coal gangue in complex working conditions based on CES-YOLO algorithm

Abstract

Aiming at the complex working conditions environmental factors such as high noise, low illumination, motion blur and mass gangue mixing in coal mines, which lead to the problems of misdetection, omission and low detection accuracy in gangue recognition, a gangue recognition model based on CFS-YOLO algorithm is proposed. The ConvNeXt V2（Convolutional Neural Network with NeXt Units Version 2）feature extraction module is adopted to replace the two C3（Cross stage partial bottle neck mudule）modules at the end of the backbone network, which effectively mitigates the feature collapse problem as well as maintains the diversity of the features in the network delivery process by adding Masked Autoencoder and Global Response Normalization layers to the ConvNeXt architecture. The Focal-EIOU (Focal and Efficient Intersection Over Union) loss function is adopted to replace the original CIOU (Computer Intersection Over Union) loss function to optimize the sample imbalance problem in the bounding box regression task by means of its Focal-Loss mechanism and adjusting the sample weights, which improves the convergence speed and localization accuracy of the model. The parameter-free attention mechanism （Simple Attention Mechanism, SimAM） is added to the back-end of each C3 module of the backbone network to enhance the model's ability of extracting key features of coal gangue targets with large scale variation or low resolution by virtue of its attention weight adaptive adjustment strategy. The effectiveness and superiority of the proposed CFS-YOLO model is verified by ablation and comparison experiments. The experimental results show that the CFS-YOLO model can effectively improve the detection effect of coal gangue under the complex environment of high noise, low illumination, motion blur and large amount of mixed coal gangue in coal mines. The mean Average Presicion (mAP) of the model reaches 90.2%, which is 3.7% higher than the mean Average Presicion (mAP) of the original YOLOv5s model, and the average detection speed reaches 90.09 FPS (Frames Per Second), which can fully satisfy the demand of real-time detection of coal gangue. Meanwhile, compared with six YOLO algorithms such as YOLOv5s, YOLOv7-tiny and YOLOv8n, the CFS-YOLO model has the strongest adaptability to the complex environment of coal mines and the best comprehensive detection performance, which can provide technical support for intelligent and efficient sorting of coal gangue.

Keywords

• yolov5s
• coal gangue recognition
• feature extraction
• loss function
• attention mechanism