A Review of Region-Based Image Retrieval

Abstract

Searching interested images based on visual properties or contents of images is a challenging problem and it has received much attention from researchers in the last 20 years. The gap between lowlevel visual features and high-level semantic understanding of images, which is also known as the semantic gap problem, is the bottleneck to further improvement of the performance of a content-based image retrieval system. In order to solve this semantic gap problem, one of the most popular approaches in recent years is to change the focus from the global content description of images into the local content description by regions (region-based image retrieval) or even the objects in images (object-based image retrieval). Although much research in region-based image retrieval has already been done, there are still three main problems need to be tackled properly: (a) local region-based features, (b) similarity measures, and (c) relevance feedback based on regions. In this paper, we review some recent development in region-based image retrieval with respect to the above three problems and propose some future directions for region-based image retrieval research towards the end of this paper.

Relevance feedback for human motion retrieval using a boosting approach

“…It is necessary for CBMR to employ effective online learning mechanisms to further refine the retrieval performance. The online learning methods can be broadly characterized as either discriminative or generative according to whether or not the distribution of the data is modelled [11]. Because discriminative methods are trained to predict the class labels rather than the detailed distribution model, they usually tend to have better predictive performance.…”

“…Because discriminative methods are trained to predict the class labels rather than the detailed distribution model, they usually tend to have better predictive performance. Support Vector Machine (SVM) and boosting are two of the most representative discriminative methods for relevance feedback [11]. SVM is considered as one of the state-of-the-art learning methods in CBIR owing to its good generalization ability [11,12].…”

“…Support Vector Machine (SVM) and boosting are two of the most representative discriminative methods for relevance feedback [11]. SVM is considered as one of the state-of-the-art learning methods in CBIR owing to its good generalization ability [11,12]. Since SVM is a kernel method, the kernel function and parameters used in SVM are very crucial in determining its performance [11,12].…”

“…SVM is considered as one of the state-of-the-art learning methods in CBIR owing to its good generalization ability [11,12]. Since SVM is a kernel method, the kernel function and parameters used in SVM are very crucial in determining its performance [11,12]. The most commonly used kernel functions such as RBF in SVM are restricted to the input space with fixed-length feature vectors, and it is suitable for CBIR.…”

Combination of Color and Significant Edge Features in Image Retrieval

“…The color representation schemes are essentially invariant under rotation and translation of the input image. Common color features or descriptors in CBIR systems include color-covariance matrix, color histogram, color moments, and color coherence vector [6]. The widely used color histograms are stable object representations in the presence of occlusion and over change in view, but it is sensitive to noisy interference such as illumination changes and quantization errors.…”