Image Classification Algorithm And Implementation Based On Feature Label Dependence

With the rapid rise of machine learning artificial intelligence,the traditional single label has been difficult to accurately describe objects,and multi-label learning has gradually become a research hotspot for everyone.In a multi-label learning framework,the more useful data,the more accurate description of the object.But with the continuous increase of information,redundant information will also increase and seriously affect the final description and judgment of the object.Feature extraction is an effective method to deal with the high dimensionality of data.By filtering some redundant or irrelevant information for the original feature space,a group of all or most of the valid information in the original feature space is selected Feature subset.But the existing multi-label feature extraction algorithms rarely make full use of feature information and fully extract "feature-label" independent information and fusion information.The feature extraction is the key points in image classification which directly affects the accuracy of classification.Based on the above feature extraction researches,we further consider the zero-shot classification of images.As we know,the image classification is always been a research hotspot in the field of computer vision,but traditional models can only classify objects that have been seen,and there is no way to handle samples without a training set.Therefore,zero-shot learning is proposed to solve unseen objects in the model.Most of the existing zeroshot learning algorithms are improvements to image feature mapping and semantic vector mapping,or use the generative adversarial network to transform the zero-shot image classification task into traditional image classification.However,none of the above methods take into account the information contained between features and labels in the original data sets.In view of the above problems,the main research work of this paper is as follows:(1)Aiming at "feature-label" and "feature-feature",a multi-label feature extraction method based on auto-encoder for feature label is proposed.This method uses a kernel limit learning machine auto-encoder to fuse the label space with the original feature space and generate it.Reconstructed feature space,combined principal component analysis and Hilbert-Schmidt independence criterion to extract "feature-feature" and "feature-label" information respectively;using a multi-label k-nearest neighbor algorithm classifier for classification.(2)Aiming at the zero-shot image classification problem,a feature-label-dependent zeroshot image classification algorithm is proposed.This method first fuses the label space with the feature space to generate a new feature space.Then,using the label information,the information loss is reduced by using principal component analysis during the extraction process,and finally the CRnet model is used for zero-shot image classification.(3)Because the data sets used in the previous experiments are all processed data,and the real image data is closer to the real world.In order to apply the theory to solve the practical problems,we participated in the third China Data Mining Competition(the first international butterfly recognition competition)and proposed a butterfly classification model.Faster-RCNN is used as the framework,the ResNet50 is used to extract depth features.The experimental results show that the experimental method can locate and classify butterfly pictures,and achieved the 21 st place in the world,which has certain practical significance.