Research On Mineral Identification Based On Multi-label Image Classification

Minerals in their natural state encounter many difficulties in classifying the discovered minerals due to their complex growth environment and difficult to control imaging.Traditional methods for classifying rocks and minerals are cumbersome,complicated,time-consuming,and inefficient.Thanks to the continuous development of deep learning,the application of convolutional neural networks has become more and more abundant.Some scholars have used this technology to solve the problem of multi-label classification of mineral image data,which has also achieved specific results.However,the research of image recognition in minerals is tough because the problem of co-occurrence of minerals in natural scenes is too complex,and the shape position of the target is uncertain.In this thesis,based on the mineral image,the correlation between the co-occurrence mineral species is used.After the operation of pre-processing the collected sample database by combining supervised learning,we start the process of feature extraction,followed by modeling the correlation between multiple labels of images and finally outputting the predicted image labels to achieve the purpose of automatic recognition and classification of mineral photos.The main work of this thesis is as follows:(1)There are unfavorable factors such as too small target,blur,and background interference in the image data set collected from the mineral site,which require preprocessing of the data,including image resolution size normalization and image data standardization,etc.The pre-processed images have higher definitions and more obvious mineral target contours,which are more suitable for experimental research.(2)A mineral recognition network for image multi-label classification based on an efficient convolutional neural network,referred to as CNN-LSTM,is proposed.Feature extraction of mineral images using convolutional neural networks,joint image labels mapping each label as an embedding vector in the joint space,combined with long and short-term memory network neurons to obtain a new output,mapping this output together with the image features through the mapping layer to the low-dimensional space to obtain a new mapping,and then path search to obtain the predicted label probabilities,with the maximum probability label as the output.(3)Three different convolutional neural networks are set up as feature extractors to compare the model advantages and disadvantages.The three models are ResNet+LSTM,EfficientNet+LSTM,and NFNet+LSTM,and according to the experimental results,the model can perform the recognition of multi-labeled images,and NFNet+LSTM is more effective.