Algorithm Research And System Design Of Intelligent Detection Of Sundries In Ore Conveyor Belt Based On Improved YOLO

There are a large amount of sundries left in the ore pile after the mine blasting,in which the manual operation takes occupied factor in the cleaning of large-scale sundries in the early stage.Limited by the environmental conditions,the operations are time-consuming,labor-intensive and inefficient,and thus some sundries enter the later stage of transportation,which lead to a impact on the beneficiation equipment and flotation process,and even causes economic losses.To the above vital issues,this thesis designs an intelligent detection system for sundries in ore conveyor belt based on deep learning algorithm.Meanwhile,we propose an improved algorithm YOLO-SD based on YOLOv3 to realize the real-time detection of sundries,which balances the detection accuracy and speed.The specific researches are as follows:(1)Detecting sundries based on the YOLOv3 algorithm.On the basis of analyzing the principles and characteristics of various detection algorithms,we determine the YOLOv3 as the benchmark algorithm for research.Firstly,expanding the data set through data enhancement methods to avoid overfitting of model training,e.g.,geometric transformation,color transformation.Then,employing Kmeans algorithm to recluster the data set to obtain a priori box that better matches the sample size.Finally,applying cosine annealing algorithm to optimize learning rate decay strategy to avoid the model falling into a local optimum.The experimental results show that the YOLOv3 presents positivity both in accuracy and speed.However,it cannot satisfy the real-time requirements on low-profile industrial platforms as the model complexity is too high.(2)Prosposing an improved algorithm YOLO-SD.By comparing the experimental results of three lightweight networks,we select the Shuffle Netv2 as the backbone network to enhance the detection speed of the algorithm,introduce the pyramid pooling module and the attention mechanism of adaptive convolution kernel size to integrate global context information and enhance feature expression capabilities.In addition,to improve the localization ability of the algorithm,we determine the CIo U as a bounding box regression loss function.The experimental results show that the detection accuracy,speed and model complexity of the YOLO-SD all meet the requirements.By the comparison with a variety of one-stage algorithms,we take a conclusion that YOLO-SD has the best comprehensive performance and stronger robustness.(3)Designing a set of ore conveyor belt sundries detection system.According to the actual needs,the overall architecture and functional modules of the sundries detection system are designed,and corresponding,the sundries detection software is developed through Python language,Py Qt5 framework and My SQL database,in which the development process includes front-end interface design and back-end logic function writing.Testing the sundries detection software based on the YOLO-SD model,and the test results show that the software realizes the functions of image display,detection model selection,detection method selection,image detection,detection result display,historical data query and other functions,which verifies the practicability of the proposed algorithm together with the feasibility of detection software.