Research On Bauxite Ore Detection Method Based On Deep Learning

Aluminum and its alloys are widely used in aviation,construction,automobile and other industrial fields because of their unique properties.Bauxite is the best raw material for generating aluminum and it's not renewable.Therefore,it is very important to make good use of the existing resources.Now,the domestic bauxite sorting still uses traditional manual sorting,which has many defects such as the inability to guarantee the safety of personnel,the slow sorting speed,and the low sorting accuracy.With the development of machine vision,the detection effect based on deep learning is far beyond the eye.Manual sorting method is difficult to distinguish samples,deep learning algorithm can also find the difference between them through a large amount of data,and the speed of sort is fast.The sorting accuracy and production efficiency of bauxite have been greatly improved,and the detection method is simple,the cost is low,so this paper uses the machine vision algorithm based on deep learning to achieve the task of bauxite sorting.The method aims at quickly and accurately realizing the identification and positioning of the four types of bauxite raw rocks with different qualities.The research is carried out from three aspects: data set production,model selection,and model training.First,the vision system and robotic arm adopt Eye-to-hand matching method,using a RGB-D camera to collect data from existing bauxite mines,and after filtering,labeling,sorting and preprocessing the collected data,it is converted into the COCO data set format required by the Detectron2 platform and the unique Darknet data of the Darknet platform,so that the experimental data set done.Second,configure the hardware environment according to the model training needs,and configure the software environment required by the GPU acceleration and target detection platform under the Linux platform,build the Detectron2 and Darknet target detection platforms,select the network model and set the parameters for the bauxite sorting characteristics,so that the experiment network model done.Finally,a model training comparative experiment was carried out,which is designed for the problems encountered in the model training of the bauxite sorting task.The datasets of the category data balance and category data unbalance were trained on Faster R-CNN and YOLOv4 models respectively,by comparing the detection effects of all the generated models,the model generated by 7000 times of data balance dataset training in yolov4 network is used as the final model of this paper.Its accuracy on test is 99%,and the inference speed is between 0.03 s and 0.05 s,which can be done in real time.Because the training process is not only a process of continuous optimization,but also a process of constantly meeting individual needs,so this paper design model optimization ideas to increase the integrity of model training.The machine vision sorting method based on deep learning designed in this paper can effectively replace manual sorting.The experience and methods are versatile and suitable for other types of ore sorting tasks and even other types of target sorting tasks.