Research On Rapid Recognition And Detection Of Coal Gangue Based On Multi-spectral Band Selection

Coal occupies a dominant position in China's energy structure.In recent years,the proportion of coal in primary energy has gradually declined,but coal consumption has still increased.This means that for quite a long time in the future,China will need to mine a large amount of coal every year,so it is necessary to strengthen the modernization of coal mines.With the rapid development of artificial intelligence and information technology,the modernization of coal mines has been put on the agenda.For example,Guiding Opinions on Accelerating the Development of Intelligent Coal Mines were issued by the State Development and Reform Commission and other departments in March 2020.Coal gangue separation is an important step in the coal mining process.Realizing intelligent coal preparation can promote the modernization of coal mines.Besides,it is also of great significance to the clean and efficient use of coal.Combine multi-spectral imaging technology with deep learning to study the recognition and positioning of coal gangue intelligent separation from experimental methods and data analysis.The research content of this paper is as follows:1.Construct a deep unsupervised learning model to obtain learning features of two-dimensional spectral images of coal and coal gangue.Improve convolution neural network classification model structure and construct A deep unsupervised feature learning model based on the convolutional neural network,denoted as 2-Dimensional Auto-encoder(2D-AE),which learns features in an end-to-end way.Only training a deep model can get the learning features of all multispectral data used in this paper.The input resolution of 2D-AE is 408×204,and the encoded learning feature-length is 64.Besides,experiments show that batch normalization can also improve the training performance of the 2D-AE model.2.Regard the mixture of coal and coal gangue as a class.Analyze the classification performance of different algorithms with different feature description operators for spectral images of different bands.CART,AdaBoost,RF,and SVM are used to identify coal,coal gangue,and their mixture.The parameters of SVM are obtained by MVO search,and the number of weak learners in AdaBoost and RF are 175 and 60,respectively.The artificial feature selections are HOG,LBP,and SIFT.The results are the average of dozens of independent repeated experiments,and the training set for each experiment is randomly divided.Experiments show that the same algorithm uses different feature descriptions,and the bands corresponding to the maximum average recognition accuracy are different.Besides,different recognition algorithms use the same feature description,and their maximum average recognition accuracy also differs.3.Combination bands selection.Select three of 25 bands as pseudo-RGB images based on the OIF and classification performance of each band.The maximum average value of OIF is 11.138,and the corresponding combined bands are 7,12,and 23.Using 2D-AE model output features,the average recognition accuracy of RF and SVM is the best in the twelfth band.Therefore,consider[7,12,23]as the optimal combination.4.Use YOLO-V4 for the identification and location of coal gangue,and its input resolution is set to 416×416,512×512,and 608×608,respectively.The results show that the larger the input resolution,the higher the average precision(AP)of coal gangue detection,but the time consumption also increases.Non-maximum suppression and confidence threshold are used to filter false positive(FP)bounding boxes that are inconsistent with the actual box,and their parameter settings are 0.4 and 0.25,respectively.When the input resolution of YOLO-V4 is set to 416,512,and 608,the average accuracy of coal gangue detection is 89.94%,91.37%,and 95.51%.Besides,YOLO-V3 and SSD are also used for coal gangue detection,and the detection result of YOLO-V4 is relatively optimal.5.Improve YOLO-V4,denoted as V4.1,to realize accurate identification and location of coal gangue.The head of the improved model adopts two kinds of scale grids,which are 17×17 and 51×51.For the low resolution of multispectral image,the input resolution of the improved model is 408x408.Experiments show that when the number of anchor boxes of V4.1 is set to 2,3,and 4,the difference in AP and time consumption of coal gangue detection is relatively small.The number of V4.1 anchor box is set to 2 for reducing the complexity of the detection model.When the V4.1 anchor box is set to 2,the coal gangue detection AP is 95.31%,and the detection time for 115 pictures is only 3.8327 seconds,which is much smaller than the 7.871 seconds of the YOLO-V4 with 608x608 input resolution.6.Design a lightweight model.To achieve faster identification and location of coal gangue,the lightweight model network is relatively shallower and narrower and uses a smaller input resolution.When the input resolution is 204×204,the detection time for 115 images is just 1.255 seconds.After filtering the predicted redundant bounding box,the detected AP of coal gangue is 91.91%,which is better than 89.94%of YOLO-V4 with 416×416 input resolution.Figure[76]table[25]reference[175]...