MULTI-USE LEARNING INSTANCE FOR OPTIMIZED IMAGE RETRIEVAL

Abstract

In the rapidly evolving landscape of digital media and artificial intelligence, the need for efficient and accurate image retrieval systems has grown significantly. Traditional machine learning approaches often require large numbers of learning instances to achieve high retrieval accuracy, which leads to substantial computational and storage overhead. To address these limitations, this project presents a novel framework titled “Multi-Use Learning Instance for Optimized Image Retrieval (MULI)”, which aims to enhance retrieval performance while minimizing the dependency on massive datasets. The proposed system introduces the concept of multi-use learning instances, where a single learning instance can be effectively reused across multiple image categories. This approach significantly reduces computational complexity and memory consumption without compromising retrieval precision. The architecture leverages a sparse autoencoder for deep feature representation, which captures discriminative characteristics of images in low-dimensional latent space. Subsequently, an improved K-Nearest Neighbour (KNN) method is employed to filter and retain the most representative multi-use learning instances, ensuring balanced data diversity. Furthermore, a multi-weight cost function is formulated to refine the retrieval decision model. Experimental evaluation on largescale image datasets comprising over 270,000 images demonstrates that the proposed approach achieves comparable or superior retrieval accuracy (AP = 0.839, AUC = 0.833) with fewer training instances compared to state-of-the-art methods.