Multi-label Classification Of Complex Scene Based On The Incremental Learning And Deep Sparse Filtering

As the multidisciplinary cross technology such as artificial intelligence, computer vision, pattern recognition, signal processing, database, and human-computer interaction technology and so on, scene classification is a research focus in the academia, it is also the key technology to solve the problem of image semantic retrieval. SAR image scene classification is different from traditional SAR image classification techniques, scene classification is not strict in pursuiting the content of the similarity between the similar images, but focus on mining intrinsic semantic information from the images by some kind of learning methods. With the development of synthetic aperture radar(SAR) technology, SAR image resolution, image content and quantity have made a lot of growth, its application is becoming more and more widely. But both scene classification method and the image retrieval technology whose object rarely involves the SAR image.Therefore, based on the above technology gap. At first, this paper proposes aimage retrieval methods using joint double anchor figure hash based SAR images. Then extendes the image database and assignes multiple class labels in one picture, and proposes a scene classification method using incremental SVM based SAR images. However, due to the incremental SVM showes the weakness in handling the problem of the data which is in high dimension and has many categories, this paper proposes a multiple label classification method based on incremental LDA and muti-label kNN. Finally, in view of the feature learning problems of multi-label data, we introduce the knowledge of deep learning,and propose a multiple label classification method based on deep integration sparse filter. In this paper, the main research results are obtained:1. In this section,we provides detailed information on a scene classification method using incremental SVM based SAR images. Including the establishment of the database of the SAR images of the million scale, and the database is labeled with a multi label; the Contourlet transform is performed on the SAR image, and the shape and texture features are obtained. Classify the feature set using incremental SVM. Evaluate algorithm performance by multi-label similarity measure function.2. In this section,we provides detailed information on a multiple label classification method based on incremental LDA and muti-label kNN, including feature dimension reduction by incremental LDA; Classify the feature set after dimension reduction using multi-label kNN; Finally, evaluate algorithm performance by multi-label similarity measure function.3. In this section,we provides detailed information on a multiple label classification method based on deep integration sparse filter, including feature learning using deep integration sparse filter algorithm; Classify learned feature set using incremental SVM; Finally, evaluate algorithm performance by multi-label similarity measure function.4. In this section,we provides detailed information on a image retrieval methods using joint double anchor figure hash based SAR images, including extract texture, shape feature based on Contourlet transform and GIST feature based on gabor filter of SAR image; Making double feature channel; Classifing the feature set by joint double anchor figure model, mapping it into the hamming space and hash coding; Evaluate the retrieval results using hamming precision.