[Paper Review] AGPCNet: Attention-Guided Pyramid Context Networks for Infrared Small Target Detection

Tianfang Zhang, Siying Cao+2 more|arXiv (Cornell University)|Nov 5, 2021

Infrared Target Detection Methodologies39 references46 citations

TL;DR

AGPCNet introduces an Attention-Guided Context architecture with multi-scale pyramid context and asymmetric fusion to achieve state-of-the-art infrared small target detection on MDFA and SIRST Aug datasets.

ABSTRACT

Infrared small target detection is an important problem in many fields such as earth observation, military reconnaissance, disaster relief, and has received widespread attention recently. This paper presents the Attention-Guided Pyramid Context Network (AGPCNet) algorithm. Its main components are an Attention-Guided Context Block (AGCB), a Context Pyramid Module (CPM), and an Asymmetric Fusion Module (AFM). AGCB divides the feature map into patches to compute local associations and uses Global Context Attention (GCA) to compute global associations between semantics, CPM integrates features from multi-scale AGCBs, and AFM integrates low-level and deep-level semantics from a feature-fusion perspective to enhance the utilization of features. The experimental results illustrate that AGPCNet has achieved new state-of-the-art performance on two available infrared small target datasets. The source codes are available at https://github.com/Tianfang-Zhang/AGPCNet.

Motivation & Objective

• Motivate accurate segmentation of small infrared targets amid clutter and noise.
• Develop a data-driven network that leverages global and local context for small targets.
• Propose a pyramid and attention-based framework to improve feature fusion and target localization.

Proposed method

• Attention-Guided Context Block (AGCB) that splits features into patches for local nonlocal-like associations and uses Global Context Attention (GCA) to model patch-level global relations.
• Context Pyramid Module (CPM) that fuses multi-scale AGCB outputs with the original feature map to form a context pyramid.
• Asymmetric Fusion Module (AFM) that adaptively fuses low-level and deep-level features with separate attention pathways to reduce feature mismatch.
• Pixel attention (PA) and channel attention (CA) are employed within AFM to enhance fusion.
• Training and evaluation on MDFA and SIRST Aug datasets with input size 256x256, using SoftIoU loss and SGD optimization.
• Ablation studies to validate backbone choices, CPM, AFM, reduction ratios, and GCA types.

Experimental results

Research questions

• RQ1 How can global and local contextual information be effectively integrated for infrared small target detection?
• RQ2 What is the impact of multi-scale context aggregation on small target segmentation performance?
• RQ3 Does asymmetric fusion of low- and deep-level semantics improve target localization and reduce false alarms?

Key findings

• AGPCNet achieves state-of-the-art performance on MDFA and SIRST Aug datasets across mIoU and F-measure.
• In ablations, deeper backbones (ResNet-34) and inclusion of CPM and AFM improve detection accuracy.
• CPM’s multi-scale design concentrates energy on target regions, reducing background interference.
• GCA with Pixel Attention (PA) and Patch-Wise attention outperforms alternatives, enhancing global–local fusion.
• AGPCNet shows strong gains on a merged dataset and demonstrates robust target segmentation with reduced false alarms.