Algae Image Retrieval Based On Multi-feature Integration And Svm Relevance Feedback

Algae is a class of eukaryotic Protista, which contribute humankind's production and life, but also are the culprits of disaster such as Algal bloom and red tide. The correct classification of algae is important for seeking the positive and avoiding the negative. Content-based image retrieval (CBIR) is kind of method which can retrieve similar images by using image indexing established by visual features and features similarity measure. In this paper, an image retrieval algorithm for the characteristics of algae was proposed to support rapid and accurate identification of algae. The main contribution of this paper was summarized as follows:1) First studied the special characteristics of algae and analyzed the limitation of traditional contour segmentation algorithm applied on algae cell image; on this basis, a new segmentation algorithm was proposed. This algorithm first remove image noise and extract color gradient, then establish a Gamma Mixture Model for color gradient histogram, after estimation of parameters in this model by using EM algorithm, solve a gradient threshold and binary the color gradient image, at last extract cell contour by means of Chain-code Method. The results of many experiments show that it could be a simple and effective algae cell contour extraction algorithm.2) The advantages and disadvantages of various feature extraction methods were analyzed and compared reciprocally, then determined the color histogram, invariant moments, Gabor transform as image color, shape, texture features represented. After image segmentation, local features of cell region were extracted. The Differentiation of local features was greatly increased compared to the overall feature. Then the organization methods of multi-feature and the establishment methods of an integrated multi-feature index for image were studied. After comparative analyzing of a variety of similarity measure methods, the initial retrieval strategy was determined.3) To further eliminate the differences between low-level features and high-level semantics, the SVM-based relevance feedback technique was introduced after the initial retrieval. To solve the problem of inadequate positive cases and negative cases, a SVM relevance feedback algorithm with memory function was proposed; then the relevance feedback process was further improved by re-using criteria of similarity measure to further improve the retrieval accuracy. 4) A retrieval algorithms experimental platform based on Matlab platform is built on. An Experiment was performed based on algae image gallery contained about 100 pictures. The results of experiments show that the results of retrieval based on integrated multi-feature were better than the ones based on single-feature search, and the results of retrieval based on relevance feedback more meet the user's requirements.