Based On The Example Of Learning Algorithms Of Image Retrieval And Recommend Related Research

With the number of digital images in Internet growing enormously, it becomes much difficult for users to effectively find what they really need from the large scale image set. Using semantic to summary, retrieve and recommend images can greatly improve the efficience of acquring images, therefore it has become an important research direction in computer vision domain.In this dissertation, some related key techniques of coarse semantic image retrieval and recommendation under the multiple instance learning framework, such as SIFT kernel construction, semantic image clustering and classification, semantic image retrieval and recommendation etc, have been explored and studied. The main work and innovations are listed as follows:(1) Focusing on the set and different location property of Image SIFT features, a novel SIFT kernel named AMK is proposed. First, the mean SIFT vector of image SIFT sets is computed, and the pair-wise SIFT points of different images are chosen based on angel similarity, next the local neighbor SIFT points of pair-wise SIFT points is considered as their context. Finally the AMK kernel is the weighted sum of the mean SIFT vector similarity, pair-wise SIFT points’similarities and context points’similarities. The AMK function is theoretically proved to be a valid kernel function and the experimental results also verify the AMK kernel’s efficience.(2) Aiming at the sensitive to noise problem of existing k-medoids algorithm named BAMIC, a new MIL clustering algorithm named ECMIL is presented. First, the similarity instances are emerged by applying the instance distance in vector space. Then adopt the Earth Mover Distance to compute the different bags’distance. Finally k-medoids algorithm is used to cluster the MIL bags. Experiments on benchmark data sets, such as MUSK、Corel and SIVAL, have shown that ECMIL can provide better clustering results.(3) Based on three assumptions of Multiple Instance Learning (positive instance clustering, bag structure and instance probabilistic influence to bag label), two MIL classification algorithms named CK_MIL and ck_MIL are presented. First, K-means clustering algorithm is applied to two sets composed from positive bags and negative bags separately, additionally select the positive instance in bags and calculate the bag structure. Then CK_MIL directly adopt RBF kernel to compute the positive instance and bag structure’s similarity, while ck_MIL introduce an probability coefficient to balance these two parts’influence to bag similarity. Finally Support Vector Machine (SVM) is used to classify the MIL bags. Experiments on MUSK dataset and image dataset have shown that these two algorithms can efficiently improve classification accuracy.(4) Focusing on object based image retrieval (OBIR), a novel image retrieval algorithm named SCAMK-MIL is proposed. The algorithm regards image as bag and segmented regions as instances. First, the positive instances of bags are acquired by using spectral clustering. Then the positive instance similarity and other instances’(excluding positive instances) similarity are computed by applying the RBF and AMK kernels. Finally Support Vector Machine (SVM) and Revelent Feedback are used to retrieval images. The experiments on SIVAL image dataset have proved SCAMK-MIL can deal well with the OBIR retrieval task.(5) The multiple instance learning (MIL) is firstly introduced to the image recommendation, and a new research direction named object image recommendation is proposed. Focusing on this problem, firstly a new combination recommendation named DD_RS based on the traditional DD algorithm of MIL and traditional cosine similarity is presented. Next, based on the color, wavelet and local feature of images, a novel CKMIL_RS recommendation framework using modified ck_MIL is provided. The experiment results on evaluation matrix of SIVAL and Caltech101have shown that these two methods can greatly improve the recommendation performance.