Efficent Sparse Reconstruction Based Hashing for Large Scale Approximate Nearest Neighbor Search

The explosive growth of data in computer vision has made Approximate Nearest Neighbor (ANN) search critical for tasks like image retrieval and object detection. Traditional tree-based methods face challenges with the curse of dimensionality, leading to inefficiencies in high-dimensional spaces. Hashing-based approaches have gained traction for their ability to optimize search time and storage. This paper presents a novel unsupervised hashing framework utilizing sparse reconstruction to enhance ANN efficiency. By integrating a structured sparsity constraint via the l2,1-norm, our method improves the discriminative power of hash codes and preserves the geometric relationships between datapoints. Unlike conventional methods that treat entropy maximization and structure preservation as separate objectives, our approach balances these goals within a unified model, ensuring hash codes are both informative and geometrically consistent. Experimental results demonstrate the effectiveness of this approach in capturing intrinsic data structures while improving performance and memory efficiency.