Infrared small target detection algorithm based on nested FPN and interference suppression

doi:10.1016/j.eswa.2025.127029

Return

Infrared small target detection algorithm based on nested FPN and interference suppression

2025-05-01

PRE

DOI：10.1016/j.eswa.2025.127029

Original

Original request for help

Save

Abstract

En 中文

Infrared small target detection (IRSTD) aims to identify small targets from complex backgrounds in infrared images. Due to pooling layers in convolutional neural networks (CNNs), which may cause the loss of small target information in deeper layers, it is difficult to transfer the traditional CNN network to the field of IRSTD. To address this problem, we propose an infrared small target detection algorithm based on nested feature pyramid network (FPN) and interference suppression loss (namely, NFPN&IS) in this paper. Specifically, the feature is extracted from input images using two sets of nested encoder blocks, which can create two mutually embedded four-layer feature pyramids. This allows the feature data of the image to flow not only from top to bottom in the feature pyramid, but also realizes data interaction in the same layer of different feature pyramids. Consequently, the proposed approach mitigates the issue of feature loss caused by excessive network depth. We further propose a loss function based on interference target suppression to achieve suppression of interference targets with similar thermal radiation characteristics of infrared small targets in complex backgrounds. Experimental results validate the capability of the proposed NFPN&IS in detecting small targets with varying sizes and low signal-to-noise ratios against complex backgrounds. They also demonstrate the suppression effect of NFPN&IS on the feature expression of interfering targets through 3D visualization. Comparison results demonstrate that the proposed method outperforms the state-of-the-art method in terms of effectiveness and robustness. The code and model are available at https://github.com/Nayutacc/NFPN-IS.

Keywords:

Convolutional neural network

Feature pyramid net

Hybrid attention

Infrared image

Small target detection

Loss function