Bilateral Cross Hashing Image Retrieval Based on Principal Component Analysis

Abstract

Image retrieval (IR) has become a crucial challenge in computer vision with the exponential growth of digital imagery. The existing methods employ a single hash source, which may overlook deep details in the image, and they struggle to handle the complexity and diversity of modern visual data. This study addresses this limitation by proposing a novel deep hashing-based IR method named bilateral cross hashing based on principal component analysis (BCHP). Bilateral cross hashing based on principal component analysis-image retrieval (BCHP-IR) employs the feature extraction capabilities of residual network-50 (ResNet-50) and the dimensionality reduction and information preservation properties of principal component analysis (PCA). The method extracts high-level features from query images using ResNet-50 and then compresses both features and class labels using PCA. The compressed data undergoes quantization to generate binary codes. These "bilateral" hash codes are combined to capture deep features and compared with image codes in the database. The BCHP-IR's effectiveness is demonstrated through extensive comparative analysis against reported methods, achieving superior performance metrics. On the MS-COCO dataset, BCHP-IR achieves mAP scores that are higher than the average of other benchmark algorithms by 6.3, 6.4, 6.2 and 5.0 at hash lengths of 16, 32, 48 and 64, respectively. These enhancements at those hash lengths are 4.6, 4.7, 4.8 and 4.3 for the NUS-WIDE dataset and 3.9, 2.9, 2.5 and 2.1 for the ImageNet dataset. Therefore, the proposed BCHP-IR method harnesses the power of ResNet-50 and PCA and offers a promising solution for efficient and effective image retrieval.