High-Resolution Feature Representation Driven Infrared Small-Dim Object Detection

Abstract

Infrared small-dim object detection is a challenging task due to the small size, weak features, lack of prominent structural information, and vulnerability to background interference. During the process of deep learning-based feature extraction, as the number of layers increases, the size of the feature map decreases, resulting in a reduction in resolution for small object features. This reduction negatively affects the network’s ability to capture fine-grained details and compromises the detection efficiency. Besides, the infrared objects can be easily overwhelmed by strong background interference, which further diminishes the original faint representations. To solve these issues, we proposes a high-resolution feature representation driven network for infrared small-dim object detection (HRFRD-Net). This network comprises three key components: High-Resolution Feature Representation Branch (HRFR), Infrared Small-Dim Object Detection Branch (ISDOD), and Spatial-Frequency Interaction Feature Enhancement Module (SFIFE). The HRFR branch employs implicit neural representation to super-resolve the infrared small objects in a self-supervised learning scheme. To effectively detect the small-scale objects, ISDOD leverages the shared encoder from HRFR to construct high-resolution and high-quality representation of infrared small objects in a resolution-free manner. To address the issue of dim objects, SFIFE incorporates a global-local mixed receptive field via the features interaction in spatial-frequency dual domains, which significantly improves the accuracy of infrared dim object detection. Experiments conducted on the MSISTD and MDvsFA datasets demonstrate the effectiveness of our approach, especially in complex scenarios where the objects are heavily obscured by the background and background interference closely resembles the objects.

Y. Dong and Y. Wang—Contribute equally to this work.