Algorithm Research On Content-based Image Retrieval

With the advent of the era of big data, image data is growing explosively. Faced with an increasing number of image data, how to effectively mine hidden information from the huge image database and quickly, accurately find the image which users really need, has gradually become a major field of computer vision area. In this case,content-based image retrieval technique was developed.This thesis concentrates on content-based image retrieval algorithms, mainly starts with feature extraction methods, taking the characteristics of image retrieval tasks into consideration, doing research on CBIR algorithms from two aspects :Algorithms based on global features and algorithms based on local features. Also,some improvements are made to the existed algorithms in order to obtain better retrieval results. The main research contents in this thesis are as follows:(1) Image retrieval based on global features. Aiming at the shortcomings of single feature image retrieval methods, put forward a multi-feature fusion retrieval method. Using Color and Edge Directivity Descriptor(CEDD) to extract color and texture features, CEDD compactly combines color and texture features, and the descriptor size is limited to 432 bit per image; Using Pyramid Histogram of Oriented Gradients(PHOG) method to extract shape features, PHOG layers the image and calculates gradient direction histogram distribution of the different layers of edge image, has some anti rotation ability and strong anti noise ability; Finally, using low Level feature fusion to combine the three features to do image retrieval task.Experimental results show that using the multi-feature fusion method can lead to better retrieval results, compared to the original methods and methods in other papers,retrieval precision was improved in varying levels.(2) Retrieval oriented image local feature extraction. Compared with the global features, local features are described in the local area of the image, which is more suitable for the retrieval of specific objects in the image. In this context, the local feature extraction algorithm is studied. Due to the characteristics of image retrieval tasks, the local feature extraction algorithm has two requirements: strong expression ability and fast extraction speed. According to these two requirements, an improved SURF algorithm is proposed. In image local feature extraction algorithm, use box filter with gradient information to form scale space, to retain more image detailsinformation; On the other hand, only calculate Haar wavelet response once in circular neighborhood of a feature point, repeated calculation is avoided, and reduce the dimension of the feature while guarantee rotational invariance. The experimental results show that the proposed method is more efficient than the traditional method,the feature points are detected more, and the feature extraction speed is obviously improved.(3) Image retrieval based on Bag of visual words(BoVW) model. Introduced the background and basic idea of BoVW model, aiming at the deficiency of classic BoVW model, a new retrieval method of BoVW model based on improved SURF feature is proposed. Firstly, using improved SURF algorithm to extract local image features; Then, using the improved k-means algorithm to cluster the feature; Finally,using tf-idf weighted method to generate weighted visual dictionary. The experimental results show that the improved BoVW model is more representative of the image,compared to the original model, retrieval precision is increased by more than 2%.This method is based on local features, compared it to the global feature retrieval method proposed in this thesis, and draws the conclusion: due to the characteristics of the global and local features, retrieval method should be chosen flexibly according to the different categories of images and retrieval requirements.