Large-Scale Image Retrieval with Deep Attentive Global Features.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Feature extraction is crucial for image retrieval, but convolutional neural networks (CNNs) struggle with clutter and occlusion.
• Existing CNN methods often yield features that lack discriminability due to interference from image complexities.

Purpose of the Study:

• To develop an attention-based mechanism for extracting highly discriminative features for image retrieval.
• To enhance the robustness and accuracy of CNNs in challenging image retrieval scenarios.

Main Methods:

• Proposed two attention modules: a spatial attention module and a channel attention module, cascaded for feature map weighting.
• Introduced a scale and mask scheme, including multiple scale filters and MAX-Mask, to refine features by addressing component scales and filtering redundant information.
• Integrated these modules into a novel network architecture for image retrieval.

Main Results:

• The spatial and channel attention modules were found to be complementary, significantly improving performance.
• The proposed network, incorporating both attention modules and the scale-mask scheme, demonstrated superior performance compared to state-of-the-art methods.
• Experiments on four benchmark datasets validated the effectiveness of the proposed approach.

Conclusions:

• The developed attention modules effectively enhance feature discriminability in CNNs for image retrieval.
• The integrated scale and mask scheme further refines features, mitigating issues related to component scale and redundancy.
• The proposed method represents a significant advancement in image retrieval technology, offering improved accuracy and robustness.