Hierarchical Visual Perception and Two-Dimensional Compressive Sensing for Effective Content-Based Color Image Retrieval

BackgroundAlthough content-based image retrieval (CBIR) has been an active research theme in the computer vision community for over two decades, there are still challenging problems in properly understanding the process in feature extraction and image matching. Consequently, significant research is still required to develop solutions for practical applications, especially in exploring and making the best using of the cognitive aspects of the human vision system.MethodologyMotivated by three cognitive properties of human vision, namely hierarchical structuring, color perception and embedded compressed sensing, we proposed a novel framework for CBIR. First, we use a hierarchical approach to perform discrete cubic partitioning of the image in the HSV space. Then, we propose a new hierarchical mapping of the image data through the use of hierarchical operators: SGLCM. These features are then integrated in a 2D CS model, which extracts refined features and suppresses noise. Finally, the resultant features are used for similarity-based ranking to perform CBIR.Results and ConclusionsExperiments were performed using two Corel image datasets, i.e., the Corel-1000 dataset which contains 1000 images in 10 image categories and the Corel-10000 dataset which contains 10000 images in 100 image categories where each category contains 100 images. In comparison with three other state-of-the-art approaches, the proposed method has demonstrated much improved retrieval accuracy, especially for images with rich color contents and detail, yet the computational complexity has been significantly reduced to meet the needs for real-time online applications. The implication of the study is that the exploitation of cognitive properties of our human vision systems in effective CBIR. Future research work can be further explored to address some limitations for optimized parameter setting, adaptive feature fusion and improved machine learning.