A Classification-Based Framework for Semantic Local and Global Image Retrieval

Abstract

This paper presents a novel framework for enhancing both local and global image retrieval by leveraging semantic region classification and deep learning. The core of our approach involves the unsupervised clustering of image regions into semantically meaningful categories. Each image in the database is subsequently represented by a membership weight vector indicating its affinity to these region categories, refined through a Convolutional Neural Network (CNN). This

representation enables a flexible and expressive query paradigm, allowing users to formulate precise queries by logically combining example regions using operators such as AND, OR, and NOT, as well as by specifying positive (example) and negative (counter-example) constraints. Experimental results demonstrate a substantial improvement in retrieval accuracy and user satisfaction compared to conventional methods, achieving up to a 20% increase in mean Average Precision (mAP) and confirming the effectiveness of our classification-based approach in bridging the semantic gap.