Research On Identification Of Mineral Sorting Based On Embedded Terminal Implementation Of Deep Learning

The clean and efficient utilization of coal resources is in line with the development concept of "lucid waters and lush mountains are invaluable assets" advocated by the state,and it is the only way for the development of the coal energy industry.The development of automatic mineral sorting technology is conducive to improving the efficiency of mineral mining,utilization,production,and processing,which also plays a good role in promoting the development and construction of "smart mines".Accordingly,this paper takes the detection of foreign matter in raw coal and mineral products as the task background,completes the detection and sorting of minerals using deep learning technology,and deplores the deep learning model to the embedded platform for the application.Firstly,the scheme design and experimental environment construction are carried out.According to the task requirements of foreign matter detection in mineral sorting,this paper designs the overall framework for the detection and identification of foreign objects in minerals,which completes the preparation of dual-energy X-ray transmission images,the selection of industrial cameras and light source equipment,and the construction of the corresponding experimental environment.Secondly,the detection and sorting of gangue foreign matter in raw coal are completed.In this paper,a method for sorting gangue by dual-energy X-ray is proposed,which consists of two stages:location and classification.In the first stage,according to the characteristics of ray transmission gray image,the detectability of feature information is enhanced through image preprocessing operations such as image quantization,binarization,and bilateral filtering denoising,and then the edge contour is segmented to separate the target foreground from the background to obtain the target and positioning information.In the second stage,based on the improved MobileNetV2 network model,the Scaled-MobileNet light-scale model is designed to classify and predict targets.The comparison experiment shows that the computation and parameter number of Scaled-MobileNet model are greatly reduced.The classification accuracy of Scaled-MobileNet model in the GPU hardware environment is up to 98.4%,and the prediction time of unit sample is 17ms.Gangue is the main foreign matter in raw coal,The detection and sorting of gangue in raw coal have been completed by the comprehensive use of the two stages.Then,the foreign objects in the needle coke of mineral products are detected and sorted.In this paper,an improved light object detection model MobileNet-YOLOv5 based on YOLOv5 is proposed,which achieves a good effect on foreign matter detection.The experimental results show that MobileNet-YOLOv5 is more lightweight and can achieve faster inference speed under the premise of keeping the accuracy basically unchanged.In the GPU hardware environment,it can achieve mAP@.5 of 0.953 and FPS of 46 frames.MobileNet-YOLOv5 has an excellent performance in foreign matter detection,which can meet the requirements of foreign matter sorting in needle coke.Finally,the hardware design and development for the embedded platform are completed.In this paper,the hardware design and AI development environment of the embedded platform are completed with Rockchip RV1126 chip as the processor.The deep learning models ScaledMobileNet and MobileNet-YOLOv5 are transplanted to the embedded terminal for application.Through the comparison of the model running results in different hardware environments,the effectiveness of the deep learning model in the embedded platform application can be verified,reflecting its practical value.This paper studies the mineral sorting and identification technology based on the deep learning embedded terminal,and designs a visual interface to apply the detection function.The mineral sorting system implemented on the embedded terminal has basically the same model inference accuracy as the PC terminal and excellent inference speed.The FPS reaches 17.8 frames when sorting foreign matter gangue in raw coal,and 9.4 frames when detecting and sorting foreign matter in needle coke,which can meet the requirements of efficient detection in industrial environments.