Research On X-ray Image Detection Method Of Tungsten Ore Based On Deep Learning

As a strategic non-ferrous metal,wolframite has many advantages,such as high density,high melting point and corrosion resistance.It has very important applications in life,industry and military.The extraction of raw ore from mines requires a series of beneficiation processes before it can become the tungsten concentrate we need.Therefore,selecting tungsten ore from the raw ore has become the most important step in beneficiation.The accuracy of ore sorting fundamentally determines the recovery rate of ore resources.With the advancement of modern digital construction,mining enterprises are gradually transitioning towards digital enterprises,However,there are still many mining beneficiation plants that use manual manual sorting for beneficiation.Workers manually select large amounts of raw ore for long periods of time every day,which can easily lead to visual fatigue for sorting personnel.Moreover,many tungsten ores are hidden inside the ore and are not easily discovered by workers,which can easily lead to human error and missed detection,resulting in a waste of tungsten ore resources,Therefore,the research on automatic detection methods for tungsten ore is a significant research direction in tungsten ore recovery and automation of personnel reduction.In this paper,based on the X-ray scanning technology and deep learning detection technology,instead of manual tungsten ore identification,two tungsten ore detection models based on classification network and target detection network are established,in view of the factors such as the large contrast difference between tungsten ore and waste rock after X-ray irradiation,the difficulty in identifying low-grade ores,the large difference in ore morphology,the large amount of ore processing and the many interferences,and combined with various model improvement schemes in the field,The network models are improved to improve the sorting accuracy of the model on the basis of ensuring the sorting rate.The main research work is as follows:1.Due to the lack of publicly available X-ray ore image classification and detection datasets in the current research field,this article constructs a self built X-ray tungsten ore image dataset,collects X-ray ore original images with a pixel size of 1664 * 256,and processes them to create a dataset.For the image classification dataset,image processing was used to cut the small ores in the original image,and a four classification and two classification dataset was established using a combination of manual and model sorting.For target detection datasets,due to the high cost of manual annotation,only a small portion of the dataset is created using manual annotation.Then,the ore image dataset is automatically generated using image generation.Finally,a partial dataset is generated using model annotation and manual confirmation.By combining these three methods,a desired target detection ore image dataset can be created at a faster speed.2.X-ray tungsten ore image detection scheme based on classification network in deep learning.According to real-time requirements,the lightweight network MobileNet V2 network is used as the backbone network,and an improvement scheme is used to improve the backbone network to improve the ore detection performance of the model.Firstly,by adjusting the expansion factor and width factor,the number of model parameters is significantly reduced,achieving the goal of model lightweight;Secondly,by embedding efficient channel attention mechanisms in some residual modules and the original model classifier,and replacing the remaining residual modules with parallel feature extraction networks containing deep hole convolutions,the model’s feature information extraction ability is enhanced and the model’s recognition accuracy is improved;Finally,use the transfer learning training method to initialize weights and accelerate model training.After improvement,the recognition accuracy of the algorithm for tungsten ore has been improved to99.10%.Compared with classic classification networks,the improved MobileNet V2 has better detection accuracy and speed for X-ray tungsten ore images,and is deployed in software systems to meet the requirements of ore sorting detection stability and real-time performance.3.X-ray tungsten image detection scheme based on object detection network in deep learning.According to the premise of recognition accuracy and recognition rate of ore sorting,Yolov5 s network is used as the backbone network,embedding attention mechanism to increase the network’s ability to extract effective features,adding EIo U loss function to increase the convergence speed of the model and the positioning accuracy of ore targets,and changing the training strategy to reduce the information loss of tungsten ore features.By improving the overall detection performance of the detection model for tungsten ore images through these improvements,it can indirectly increase the tungsten ore recovery rate.The experiment shows that the improved Yolov5 s has a detection accuracy of 97.63% and a recall rate of 97.05% for tungsten ore,and the model has a detection rate of 41.11 frames per second for ore images.Compared to other classic object detection algorithms,the improved Yolov5 s algorithm has better performance and is deployed in software systems to meet the detection stability and real-time requirements of ore sorting.