Article: Single-stage landmark retrieval with texture feature fusion Journal: International Journal of Information and Communication Technology (IJICT) 2024 Vol.25 No.9 pp.43 - 59 Abstract: Existing landmark retrieval models typically fuse global and local feature descriptors of target images to generate feature vectors for landmark retrieval. However, these models often exhibit poor resilience to complex viewpoints, occlusions, and lighting conditions. Moreover, the fused feature descriptors still contain substantial redundant information, leading to decreased retrieval accuracy. To address these issues, this paper proposes a novel single-stage image retrieval model enhanced by texture augmentation. The model incorporates a texture enhancement module that leverages texture feature encoding to reconstruct the original feature maps, amplifying the influence of texture features in deep feature vectors across different scales. This approach ensures robust feature representation under extreme angles, occlusions, or varying lighting conditions. To mitigate the problem of redundant features, the model introduces an innovative feature fusion module. This module optimises local features from multi-scale feature descriptors using a mapping fusion technique, eliminating redundant information and generating more compact and discriminative feature descriptors. Extensive experiments demonstrate that the proposed model achieves significant improvements in retrieval performance compared to state-of-the-art image retrieval models, while maintaining acceptable retrieval times. Inderscience Publishers - linking academia, business and industry through research

Title: Single-stage landmark retrieval with texture feature fusion

Authors: Kun Tong; GuoXin Tan

Addresses: National Research Center of Cultural Industries, Central China Normal University, WuHan, 430079, China ' National Research Center of Cultural Industries, Central China Normal University, WuHan, 430079, China

Abstract: Existing landmark retrieval models typically fuse global and local feature descriptors of target images to generate feature vectors for landmark retrieval. However, these models often exhibit poor resilience to complex viewpoints, occlusions, and lighting conditions. Moreover, the fused feature descriptors still contain substantial redundant information, leading to decreased retrieval accuracy. To address these issues, this paper proposes a novel single-stage image retrieval model enhanced by texture augmentation. The model incorporates a texture enhancement module that leverages texture feature encoding to reconstruct the original feature maps, amplifying the influence of texture features in deep feature vectors across different scales. This approach ensures robust feature representation under extreme angles, occlusions, or varying lighting conditions. To mitigate the problem of redundant features, the model introduces an innovative feature fusion module. This module optimises local features from multi-scale feature descriptors using a mapping fusion technique, eliminating redundant information and generating more compact and discriminative feature descriptors. Extensive experiments demonstrate that the proposed model achieves significant improvements in retrieval performance compared to state-of-the-art image retrieval models, while maintaining acceptable retrieval times.

Keywords: landmark retrieval; feature fusion; multi-scale fusion; feature enhancement.

DOI: 10.1504/IJICT.2024.143328

International Journal of Information and Communication Technology, 2024 Vol.25 No.9, pp.43 - 59

Received: 30 Jul 2024
Accepted: 20 Sep 2024
Published online: 13 Dec 2024 *

Title: Single-stage landmark retrieval with texture feature fusion

Keep up-to-date