DESIGNING LOCAL PATTERNS FOR CONTENT-BASED IMAGE RETRIEVAL AND CLASSIFICATION USING MACHINE LEARNING TECHNIQUES

Abstract

Retrieving similar images from a dataset of images is always challenging for researchers, and it becomes more difficult under critical conditions like illumination variation and different facial expressions. Every image comprises three types of content: color, shape, and texture. The texture content of the image plays a vital role in the image retrieval process. Many researchers have been working on the image retrieval problem and have proposed many local descriptors for the last two decades. This work aims to achieve a comprehensive advancement in content-based image retrieval by addressing several objectives. First, it conducts an extensive comparative analysis of existing Local Binary Pattern (LBP) variants. Next, the traditional Local binary pattern (LBP) compares the central pixel with all 8-neighboring pixels in a 3 × 3-pixel window to generate LBP codes. However, its circular structure may produce similar LBP codes for different structural patterns. This study endeavors to pioneer a novel Hyperbolic local binary pattern (HLBP) that follows the hyperbolic structure to extract the discriminative features. By including the mean of the magnitude difference between each neighboring pixel and the center pixel, the study investigates the creation of hybrid and extended LBP variations, expanding the scope of existing research. In recognition of the important aspect of dimensionality reduction, the study also looks into integrating Principal Component Analysis (PCA) to simplify the feature space. Machine learning techniques are used to classify images using the AT&T, Yalefaces, and YaleB facial image datasets. Finally, to ensure the correctness and reliability of the results, this study uses the Kruskal-Wallis test, a strong statistical validation technique. These contributions offer valuable insights for the larger image processing and pattern recognition domains and have the potential to have a major impact on content-based face image retrieval. The major objectives of the present work are as follows: : • Background of local patterns and comprehensive comparison of existing LBP variants for content-based image retrieval Design & development of novel local binary pattern variants for gray-scaled and RGB images. • Hybrid/Extended LBP variant development for content-based image retrieval using different datasets. • Feature dimensionality reduction using PCA and classification using machine learning techniques. Objective 1: Background of local patterns and comprehensive comparison of existing LBP variants for content-based image retrieval. This objective primarily takes the form of a comprehensive survey paper, shedding light on the strengths and limitations of various LBP approaches. Contributions: The key contributions are as follows: • Background of CBIR and the need for Local patterns. • Next, we briefly discussed and compared all the existing local descriptors LBP, MBP, 6 × 6 MB-LBP, CSLBP, LTP, t-LBP, CLBP, HELBP, VELBP, ELBP, NILBP, RDLBP, LDGP, XCSLBP, LTriDP, MRELBP-NI, LCLBP, CSLBP/Centre, LNDP, LTrDP, NCDB-LBPac, NCDBLBPc, ODLBP, ELD, and RLBP. Objective 2: Design and development of novel local binary pattern variants for gray-scaled and RGB images. The second objective of this research is to conceive and design innovative Local Binary Pattern (LBP) variants tailored for gray-scaled and RGB images. This entails the creation of LBP variants with distinct features and capabilities. Contributions: The key contributions are as follows: • Proposed HLBP, a new CBIR feature descriptor, following horizontal and vertical hyperbolic movements for ‘robust’ feature extraction. • A detailed CBIR analysis on five publicly available databases is performed using the proposed HLBP descriptor, and its performance is compared with state-of-the-art feature descriptors. • Also, two more descriptors, RTLBP and MVM-LBP, are developed, and results are discussed. Objective 3: Hybrid/Extended LBP variant development for content-based image retrieval using different datasets. As an integral part of this study, the second objective involves the development of hybrid or extended LBP variants. These variants will be constructed by integrating and fusing existing descriptors, enriching the descriptor landscape in image analysis. Contributions: The key contributions are as follows: • Our first contribution is the usage of the mean of the neighborhood’s absolute difference from the center in both ETLBP and ERDLBP algorithms. • Next is the computation of the neighborhood’s absolute difference from the center in the clockwise and radial difference of ETLBP and ERDLBP, respectively. • Finally, a detailed analysis of the proposed descriptors is performed on four publicly available databases. Objective 4: Feature dimensionality reduction using PCA and classification using machine learning techniques. The fourth and final objective revolves around feature dimensionality reduction, a pivotal step in the image analysis. This objective focuses on overcoming the challenges of feature selection after feature extraction, employing Principal Component Analysis (PCA) as a reduction technique. Contributions: The key contributions are as follows: • Proposed a novel local pattern MELBP utilizing a 5 × 5 pixel window and generating two patterns, MHELBP and MVELBP. • The proposed approach is tested on three benchmark databases, which include AT&T facial images, Yalefaces, and YaleB databases, with and without PCA. The proposed method achieves very encouraging outcomes. • Classification accuracy of various machine learning algorithms is computed using local pattern features.