Research On Coal Gangue Identification Based On Deep Learning

With the continuous improvement demands of cleaning and intellectualization of coal gangue separation,the concept of automatic separation of coal and gangue was put forward by domestic researchers.One of the technical difficulties to realize the automatic separation of coal and gangue is to accurately complete the recognition and localization of gangue images.At present,some scholars have carried out extensive research in the aspect of coal gangue target recognition,but the traditional method of coal gangue recognition has some problems such as difficulty in feature extraction,weak generalization ability and unrealized localization.Aiming at the above problems,a method of coal gangue recognition based on deep learning is proposed in this paper,in order to improve the accuracy of coal gangue recognition and localization.The coal and meteorite images are analyzed and the actual working conditions are determined.The coal and meteorite image sample collection principle is determined,and the collected coal and meteorite images are screened to ensure the validity of the coal and meteorite image sample sets.The size and image training is used,and the coal and meteorite image data sets are expanded by methods such as random cropping,image flipping,brightness change,and adding noise.The coal and meteorite image datasets were re-divided using leave-one cross-validation to lay the data foundation for coal and meteorite image recognition.In the process of image recognition,the convolutional neural network has different network structure and different image recognition effects.The effects of three convolutional neural networks,AlexNet,VGGNet and GoogLenet,on the identification of coal and vermiculite are compared.Fitting the phenomenon,adding regularization,weight attenuation factor,optimizing the network model,and conducting coal and gangue image Identification experiments.In order to improve the recognition rate of the model,a migration learning based on convolutional neural network is proposed.The data normalization is processed in the network by batch normalization algorithm.The noise is added to the convolutional neural network by using the DisturbLable algorithm to optimize the network model.Improve model recognition.using the DisturbLable algorithm to optimize the network model.Improve model recognition.The experimentally verified improved model has improved recognition on coal and meteorite datasets.