Semi-Supervised Content-Based Image Retrieval

With the development of technology and the improvement of people’s life, the image retrieval is indispensible in people’s life. Most existing retrieval systems (i.e. Bing, Google, Baidu) are text-based, but the precision of the search result is largely limited by the mismatch between the true relevance of an image and its relevance inferred from the associated textual descriptions. This problem prompted increasing interests in the development of content-based image retrieval (CBIR). However, the single feature might result in the poor performance because of inappropriate metrics for descriptor comparison; noisy descriptors; feature detection drop-out; loss due to descriptor quantization and the semantic gap.This paper discussed the different low-level visualfeature and the reranking methods. With introduction of the hot topic "deep learning", our approach can nonlinearly fuse the global and local feature based on autoencoder method, which enhance the complementary ability of visual feature. Meanwhile, inspired by the spirit of query expansion, we present a semi-supervised reranking method. Concretely, we cluster and index the dataset with the help of the limited labeled data offline. In the online stage, we detect these confident samples as training data, and thentrain RankNet as a re-score kernel function by Accelerated Mini-Batch Stochastic Dual Coordinate Ascent (ASDCA), which improves the retrieval performance.In the experiment, our approach has been successfully used in the Paris, Francelandmark and TRECVID INS2012(Instance Search). It proved that our method is able to refine different initial retrieval list (low precision and high precision), which means that the final results would not be affected by the initial rank. Meanwhile, the semi-supervised model not only reduces the computation complexity but also the semantic gap.