Scene Classification Of Remote Sensing Image Based On Multipath And Hierarchical Sparse Coding

In the era of the rapid development of the multimedia information technology, more and more digital images have appeared. How to search the images that the users are interested in or classify the images for afterward data processing is a very pressing task.Image scene classification is a kind of technology that obtains the category of image automatically according to the image content. This technology has widely used in pattern recognition and computer vision. Scene categorization of remote sensing image is an important branch of image scene classification. In recent years, it has made a great contribution to the research of target detection, image retrieval, image enhancement and so on.Scene categorization of remote sensing image needs to extract features from the image firstly, and then selects an appropriate classifier to classify. So image feature extraction is very important. A good classifier will also fail for bad features. The methods of scene categorization of remote sensing image mainly divide into two categories: the methods based on the low-level features and the methods based on the middle-level features.Because the methods based on the low-level features do not need to identify the specific objects in the scene image, the computational complexity is relatively low. But for complex scene images, the methods based on the low-level features are disabled. This is the semantic gap between low-level features and high-level semantic messages. To overcome the semantic gap, the methods based on the middle-level features appears.These methods build the bridge between the low-level features and the high-level features. Our paper has three improvements on scene categorization of remote sensing image:1. We introduce a semi-supervised method of scene categorization of remote sensing image based on the multipath and hierarchical orthogonal matching pursuit(OMP).Different from the traditional methods based on the feature descriptors, this method learns the codebook from the original image blocks directly, and uses the sparse coding algorithm of orthogonal matching pursuit(OMP) and spatial pyramid matching(SPM)to obtain the representation of image, then combines the idea of multipath and hierarchical learning and max-pooling to build the unsupervised learning frameworkwhich is based on image blocks of different size. Finally, we use the Semi-Supervised Support Vector Machine(S3VM) to classify. We also apply this method to the scene detection of remote sensing image. The experimental results show that this algorithm has a good performance on the scene categorization and scene detection of remote sensing image.2. We propose a method of scene categorization of remote sensing image based on local feature descriptors and hierarchical sparse coding. This method changes the traditional learning mode of Bag of Features(Bo F) model which is a process of single scale and single layer learning. We extract the local feature descriptors of Scale invariant feature transform(SIFT) and Local Binary Patterns(LBP) on multiple scales of image blocks firstly, and then use the hierarchical sparse coding respectively on different scales of image blocks. Finally, we connect all the image feature vectors from different scales and use the Support Vector Machine(SVM) to classify. Compared with the traditional algorithms based on the feature descriptors of SIFT and LBP, the accuracy of our method is improved a lot.3. Based on the platform of MATLAB parallel computing, we design the parallel algorithm of the scene categorization of remote sensing image based on the multipath and hierarchical orthogonal matching pursuit(OMP). The operations of dense sampling,encoding, pooling in the original algorithm are all the process that the same algorithms handle different data independently, and the large amount of data make the process of parameter optimization very difficult. Using the MATLAB parallel toolbox, we design these independent calculate procedure to execute parallelly. The comparative experiments show that parallel algorithm we design reduces the time complexity greatly and solves the problem in the process of parameters optimization.