Research On Detection Method Of Foreign Object On Coal Conveyor Belt Based On Computer Vision

The belt conveyor is the key equipment for underground coal transportation.Noncoal foreign bodies such as gangue and iron(anchor,angle steel)enter the coal belt system,which may cause serious accidents such as scratches and tears on the conveyor belt,which will affect the safe production of coal mines,even cause very heavy economic losses.Aiming at this problem,this thesis mainly studies the foreign matter recognition method of coal conveyor belt based on machine vision,analyzes and summarizes the image enhancement method and the operation process of the deep learning target detection framework Faster-RCNN,and combines the characteristics of mine belt monitoring image to study the downhole image based on Retinex.Enhance methods and improve related processing in Faster-RCNN.Firstly,the author understands the theoretical background and basic ideas of Retinex image enhancement algorithm,analyzes and summarizes the classic path-based Retinex algorithm and the center/surround-based Retinex algorithm from the aspects of implementation and algorithm performance,and analyzes the characteristics of the monitoring image of the downhole belt conveyor.Based on the Retinex algorithm,the downhole image enhancement method is used to obtain the illuminance component and the reflection component of the image.A new S-curve model is used to homogenize the image illuminance distribution,and the image contrast is CLAHE adaptively balanced.The component and reflection components acquire the final enhanced image.Experiments show that the proposed method based on Retinex algorithm can better balance image brightness,improve image detail display and improve image quality.The image enhancement algorithm provides a good data foundation for the foreign object detection model.Analyze the process of image feature utilization,RPN candidate frame generation and candidate frame screening in the deep learning target detection framework Faster-RCNN.For the problem of multi-scale target detection,based on the analysis of existing image feature utilization framework,propose a bidirectional feature pyramid.The network(DSFPN)implements a new multi-scale fusion of image features using top-down and bottom-up processes.In the Faster-RCNN image feature map,about 20 K anchor candidate frames are generated by RPN.How to screen out the candidate frame with strong target character has a great influence on the efficiency and accuracy of target detection.For this problem,the original Faster is analyzed and summarized.A candidate box screening method for non-maximum suppression and hard threshold in RCNN,proposed candidate frame screening method(PSIF)for joint IoU and feature difference analysis,and verified by multiple sets of experiments,the FRCNN+DSFPN+PSIF target proposed in this thesis The detection frame can effectively improve the accuracy of target detection,and has certain applicability and application value in the identification of foreign objects in coal conveyor belt.