The Applications Of Adaptive Deep Learning Algorithm In Object Classification

Pattern classification has always been the core problem in the field of computer vision and machine learning. It is also an important research direction for domestic and overseas researchers. Because there are many applications in the direction, such as image retrieval, object classification and recognition. And feature extraction is the most important link in the pattern classification, so it has important theoretical and practical significance on the study of feature extraction. The design of a good feature extractor can not only ease the people from heavy labor in terms of features, and can also improve the accuracy rate of classification results. In the background of object classification, based on the framework of deep learning, a feature extractor of depth adaptive is put forward. And next the standard features of the image are fused with the deep features Based on SVM classifier, a weight classifier about feature fusion is proposed. The main contents of this thesis are as follows:(1) The design of depth structure, including the number of nodes in each layer and the depth of the structure. In this thesis, an algorithm of the layers self-adjustment is proposed. Compared to the original layer design method, the algorithm has better robustness. For classification problems in general, it can be able to extract better features from the original sample.(2) Standard feature extraction and selection of red bean samples. Because the characteristics of red bean samples and impurities, the features of the color, shape, texture are analyzed. Finally, a conclusion is drew that the feature of color is good to separate all kinds of samples. Shape is bad for the qualified and unqualified samples, but it is good for beans and impurities. The texture feature is bad for the qualified and unqualified samples, but it is good for beans and impurities. At the same time, if a qualified sample is classified into unqualified samples, it may have relatively small impact. But if the impurities are classified to qualified samples, it may bring very bad consequences. In order to reduce the risk, according to the consequence of classification may bring, a strategy of weight update is proposed.(3) In the part of feature fusion and classifier selection, an algorithm of feature fusion is proposed, which can minimize the risk of classification consequences. According to the SVM training model we get from the features, we can get four classification results from the models we have trained and multiplied by the weight of each feature, we can get the confidence of which class the sample belong to. Experiments show that, classification accuracy according to the algorithm proposed in this thesis is obviously better than the direct feature fusion, and it can also significantly reduce the loss brought by misclassification.Based on the analysis and research on feature extraction, the thesis focus on the work of the influence of features learned by deep learning and different feature fusion method. In the framework of deep learning, an algorithm of feature fusion is proposed, and at last a classification system is designed.