Semi-supervised Image Classification Based On Electrical Networks

The image classification is an image processing method which distinguish the image targets into different categories. It is the basic of image analysis and target recognition, and it occupies an important position in the image engineering field. The graph-based semi-supervised learning method establish the classification model from a small amount of labeled samples and a large number of unlabeled samples. Because of the advantages such as optimizing easily, it has been used more and more in the image classification. The electric network method is one of the graph-based semi-supervised classification algorithm. However, there are some problems in its practical application. Firstly, the method uses all the pixels in the image to build the graph model. As a result, it will lead a high computational complexity and impact on the classification effectiveness. Secondly, it can't classify the large image because of the way of building graph model. It seriously limits its application in image classification. It is very necessary to solve the problem of reducing the graph scales. For improving the effectiveness of the algorithm, in this passage, we consider to take advantage of the clustering method to preprocess the original samples.Based on analyzing the latest research of the electrical network system theory and the exist problems, we research the electric network learning theory, algorithms and their application in the image classification. The main research content in this paper is partitioned into the following two points:1.We proposed image classification with p-voltages algorithm based on mean shift. The original p-voltages algorithm is a semi-supervised classification algorithm, which based on samples theoretical voltages in the electricity network system. There are some disadvantages of the p-voltages algorithm for image classification, such as all the samples are used to establish a complete graph easily causes the system out of memory when applying to image classification problem, and only two labeled samples (source and sink node) are selected, resulting in poor accuracy of classification. In allusion to the above problems, image classification with p-voltages algorithm based on mean shift was proposed. This method combined p-voltages classification algorithm with mean shift algorithm, the smoothing image after mean shift algorithm as objects to be classified, for reducing the diversity of image features. And choose multi-labeled samples from the smoothed image as multiple source and sink nodes to improve the effectiveness of learning. A sample in each smoothed area was randomly selected as unlabeled sample to ensure that it carried abundant image feature information. In order to reduce the scale of composition and furthermore provide the conditions for mass image classification, the labeled an unlabeled samples are used as the subset of the original image to build graph. The experiment results indicate that the proposed method not only improves the classification accuracy, but also received a higher time efficiency, suitable for large-scale complex features image classification.2.We proposed the image classification with p-voltages algorithm based on the parameter optimization. For the binary classification problem, in the foreground and background, it would lead that the pixels belong to different category have the same color or the pixels belong to the same category have different colors in different areas due to the own color or the light. The image classification with p-voltages algorithm based on mean shift is based on the image color features to classify the image. When classifying the images with complex characteristics and fuzzy boundaries, the classification accuracy will be lower. In allusion to the above problem, we propose the image classification with p-voltages algorithm based on parameter optimization.By optimizing the threshold voltage which is an important parameter in the p-voltage algorithm, the method reduce the wrong classification and improve the classification accuracy compared with the original image classification with p-voltages algorithm based on mean shift. Otherwise, we applied the improved method to the multi-spectral remote sensing images. The experimental results show that the improved method is not only applicable to classify the ordinary image with complex characteristics, but also applies to the multi-spectral remote sensing image classification.