Multi-object detection of iron foreign bodies in scraper conveyor based on improved Mask R-CNN

The scraper conveyor is the key transportation equipment in the coal mine. The iron foreign body entering the scraper conveyor will lead to wear and tear, chain breakage, and even cause serious accidents such as production stoppage and personal injury. The existing scraper conveyor foreign bodies identification method has the problems of poor adaptability to underground images and the incapability of distinguishing the types and quantities of foreign bodies. To solve the above problems, a multi-object detection method for iron foreign bodies in scraper conveyor based on improved mask region-convolutional neural network (Mask R-CNN) is proposed. The image enhancement algorithm based on the Laplace operator is used to preprocess the images collected under the environment of low illumination and high dust. The enhanced images are marked to make a data set. The ResNet-50 feature extractor of the Mask R-CNN model is used to obtain the image features of iron foreign bodies. The feature pyramid network is used for feature fusion to ensure both high-level semantic features (such as category, attribute, etc.) and low-level contour features (such as color, contour, texture, etc.), so as to improve the accuracy of small-scale iron foreign body identification. To solve the problem that the anchor point generated by the Mask R-CNN model does not correspond to the size of the iron foreign body to be detected, the Mask R-CNN model is improved. K-means Ⅱ clustering algorithm is used to replace the original anchor point generation scheme. The cluster center point is obtained by traversing the length and width information of the tag box in the data set, so as to achieve the multi-object detection of iron foreign bodies in the scraper conveyor. The experimental results show that the average detection time of the improved Mask R-CNN model is 0.732 s, which is shortened by 0.093 s and 0.002 s compared with Mask R-CNN and YOLOv5 respectively. The average precision is 91.7%, which is 11.4% and 2.9% higher than that of Mask R-CNN and YOLOv5 respectively.