Harnessing Deep Features for Improved Multi-Query Texture Retrieval

Abstract

Developing an efficient classifier-based image retrieval system is vital for accurately and swiftly retrieving relevant images in computer vision applications. Hand-crafted features usually require extensive tuning and may fail to generalize across different types of images, making the retrieval process labor-intensive and less adaptable. Despite the advancements in deep learning for image retrieval, there is limited research on integrating Multi-Query (MQ) techniques with deep features for image retrieval. The novel MQ Deep Image Retrieval (MQDIR) system exploits this approach to extract deep features from an Image Set (IS) and handle MQ simultaneously. The methodology enhances the retrieval process by capturing more nuanced image characteristics through using MQs that traditional methods might overlook. A new precision-based metric is introduced in this study to offer a comprehensive average performance evaluation. The metric considers the precision of retrieval results across multiple ISs and Convolutional Neural Networks CNNs and allows a finer assessment of system performance compared to conventional measures. The experiments are conducted on popular benchmark ISs, including texture images, and demonstrate that MQDIR consistently outperforms existing methods in terms of retrieval accuracy and efficiency.