<p>Hashing algorithms are pivotal in large-scale approximate nearest-neighbor searches due to their rapid execution and user-friendly attributes. Recent research underscores the superiority of deep-supervised hashing algorithms over less intricate non-deep-supervised counterparts. These sophisticated algorithms leverage deep learning models to generate hash codes through nonlinear transformations and robust feature extraction, primarily for efficient image retrieval. However, the computational demands and memory requirements of many hashing algorithms are substantial. Compact deep learning models using knowledge distillation (KD) have emerged to address this challenge to alleviate computational burdens. An inherent challenge within this framework is selecting an optimal KD strategy. This paper introduces an innovative approach, “multiscale object-driven knowledge distillation” (MOD-KD), which is designed to enhance optimal deep supervised hashing for image retrieval. MOD-KD enhances student network training by addressing background interference. This involves identifying object locations in training images and segmenting inner layers of teacher and student networks into blocks. This approach preserves object features while discarding background elements, thereby improving student network training quality without misinformation propagation. MOD-KD also improves teacher network training. This approach refines KD by removing background feature interference and utilizing intermediate layer information. Extensive evaluation of CIFAR-10, CIFAR-100, and NUS-WIDE datasets demonstrates MOD-KD’s superiority over existing deep hashing methods.</p>