Research On Foreign Objects Recognition Of Coal Transport Belt Based On Deep Learning

Belt conveyors are one of the key equipment for transportation in coal mines.Large pieces of coal gangue and non-coal foreign objects such as bolts,drills,steel bars,etc.will produce serious accidents such as belt scratches,tearing and jamming when that enter the coal belt.Affectes the normal production of coal enterprises and cause serious economic losses at the same time.Based on these circumstances,we propose a deep learning-based coal mine belt transportation foreign objects identification method using the monitoring camera which already deployed in the coal mine.At the same time,we also analyze the image enhancement method and the anchor-free objects detection method with the characteristics of coal mine belt transportation foreign objects detection.The image enhancement method based on deep learning zero-reference depth curve estimation is studied,and the objects detection method based on Center Net is improved.Finally,a visualization system is designed to realize the integration of monitoring and foreign objects detection.Firstly,we analyzed the typically image enhancement method based on color histogram and light propagation model,and introduced the Enliht GAN enhancement method based on deep learning using paired non-reference data,and the superiority and drawbacks of its enhancement performance are analyzed and compared.The coal mine conveyors have faster running speed and the images which are obtained from monitors directly usually will be darker due to weak light.We suggest a image enhancement method which is called ZR-DCE based deep learning.We only train the network by light brightness which are obtained from coal mine directly,using the loss function which use a linear combine spatial consistency loss,exposure control loss,color stability loss and lighting smoothness loss to optimal it.Finally,we get a matrix by estimating parameters of deep curve.We can acquire enhancement image using the matrix multiply with original image,and the results better when the num of iterations are bigger.Then we filter the noise which generated from the original images and during the enhancement by gaussian filters.The results show that ours images enhancement method can not only increase the brightness significantly,but also retain the contrast among the near pixels and highlight the details between different objects.The method makes the quality of original input images improved significantly.The enhancement processing time can ignore for objects detection as it only take one millisecond for one frame.he enhanced image provides a good data basis for the objects detection algorithm and improves the accuracy of objects detection.It briefly discusses the development process of objects detection,and points out the advantages and shortages of anchor-based and anchor-free objects detection methods,and we introduce the features and processing of three anchor-free objects detection methods which are FCOS,Corner Net(-Lite)and Center Net-Triplet.For the complexity and real-time requirements of foreign objects detection in underground coal mines,a center-based objects detection method which convert objects detection to keypoints detection problem is proposed,the method detects the center of the object in the image,and regress to size of the bounding box.The following improvements have been made in this paper to address the characteristics of foreign objects detection in belt transport.Firstly,the Hourglass which is more reliable for keypoints detection was chosen as the backbone network.Then using deep separable convolution in residual modules instead of standard convolution,the number of parameters decrease greatly,and the running time of training and inference stage cut down significantly.At the same time,we propose group normalization to reduce the loss during training of network,which can address stability degradation from training with small batch size samples due to hardware constraints.Finally,We suggest a feature fusion method with weights that allows the features of each stage of the backbone to be fully utilized,which can solve the problem that images have complex backgrounds and heavy disturbing with coal.The output features maps have the richest features and the accuracy is improved greatly for foreign objects detection in belts.Experimental results show that the detection accuracy of the proposed method for coal gangue and iron are respectively 98.4% and95.5%.And the detection speed is about 24 fps,which can satisfy the needs of accuracy and real time.It has high actual practice in the detection of foreign objects in coal conveyor belt.A deep learning based foreign objects detection system for coal mine belt conveyors is designed based on the above research.The system can monitor four cameras simultaneously.It also features live screen viewing,enhanced screen and monitoring screen viewing,alarm message display,viewing of historical alarm messages and camera configuration.The paper has 59 pictures,7 tables,and 96 references.