Research On Coal And Gangue Target Recognition Algorithm For Gross Coal On The Ground

In order to promote the development of clean coal technology in the coal industry,gangue sorting has become an important means of efficiently improving coal quality,which can reduce the ash and sulphur content of raw coal,while gangue identification is the key to gangue sorting,and the target location and category provided by gangue identification can provide action information for subsequent sorting equipment.The current gangue identification method is more mature is the ray identification method,but the ray is harmful to the environment at the same time,it is difficult to locate the gangue target through a single-dimensional detection results;image identification method due to its green,safe and other characteristics,has become a research hotspot in the field,but the current stage of gangue image identification are used to learn the classification method of representation,such problems as machine learning in the classification task,there is recognition target The problem of single,difficult to meet the actual needs.Based on this bases,the thesis proposes the research of ground gross coal and gangue target recognition algorithm to achieve the recognition of the coal and gangue target on the belt.In this paper,the coal and gangue target on the belt of coal building is taken as the research object in the context of a coal mine of Ping Coal Group.Firstly,by using manual extraction and convolutional neural network to extract features,the features extracted in different ways are clustered with the features output from different convolutional layers through K-means,so as to propose the depth of convolutional neural network for extracting coal and gangue image features;secondly,the coal and gangue data of attached coal mud is increased through style migration model and affine transformation to expand the original data set and improve the accuracy of attached coal mud Again,the semi-automatic labeling algorithm based on semi-supervised learning is proposed for the problems of redundant labeling and excessive human resource input of the gangue image;secondly,the scale is proposed for the 50-150mm gangue target detection model scale by using the correspondence between scale and perceptual field,and the clustering algorithm is used for the target detection model according to the target size information in the gangue dataset.The anchor frame provides an a priori frame,and for the gangue detection is an intensive detection problem,an improved training loss function is proposed to obtain a target detection model suitable for gangue target detection.Finally,the mean accuracy,detection speed,and training loss of different scales,anchor frames,and confidence levels are discussed by the greedy algorithm under the same hardware loading and software deployment,and the learning rate preheat+learning rate decay is used to fully train the gangue data features.The experimental analysis of the thesis verifies the reliability of the proposed gangue target detection model,and the algorithm structure applicable to gangue target detection is given in order to adapt to the speed and accuracy requirements of gangue detection in the field,which better realises the need for gangue target detection in coal buildings.