LR-Net: A Lightweight and Robust Network for Infrared Small Target Detection

Chuang Yu,Yunpeng Liu,Jinmiao Zhao,Zelin Shi

2024-08-06

Abstract:Limited by equipment limitations and the lack of target intrinsic features, existing infrared small target detection methods have difficulty meeting actual comprehensive performance requirements. Therefore, we propose an innovative lightweight and robust network (LR-Net), which abandons the complex structure and achieves an effective balance between detection accuracy and resource consumption. Specifically, to ensure the lightweight and robustness, on the one hand, we construct a lightweight feature extraction attention (LFEA) module, which can fully extract target features and strengthen information interaction across channels. On the other hand, we construct a simple refined feature transfer (RFT) module. Compared with direct cross-layer connections, the RFT module can improve the network's feature refinement extraction capability with little resource consumption. Meanwhile, to solve the problem of small target loss in high-level feature maps, on the one hand, we propose a low-level feature distribution (LFD) strategy to use low-level features to supplement the information of high-level features. On the other hand, we introduce an efficient simplified bilinear interpolation attention module (SBAM) to promote the guidance constraints of low-level features on high-level features and the fusion of the two. In addition, We abandon the traditional resizing method and adopt a new training and inference cropping strategy, which is more robust to datasets with multi-scale samples. Extensive experimental results show that our LR-Net achieves state-of-the-art (SOTA) performance. Notably, on the basis of the proposed LR-Net, we achieve 3rd place in the "ICPR 2024 Resource-Limited Infrared Small Target Detection Challenge Track 2: Lightweight Infrared Small Target Detection".

Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

The problem that this paper attempts to solve is that the existing infrared small - target detection methods are difficult to meet the comprehensive performance requirements in practical applications. Specifically, due to device limitations and the lack of inherent characteristics of targets, it is difficult for the existing infrared small - target detection methods to achieve an effective balance between detection accuracy and resource consumption. Therefore, the author proposes an innovative lightweight and robust network (LR - Net), aiming to solve the following problems:

1. **Balance between detection accuracy and resource consumption**: Existing methods are often too complex, resulting in excessive resource consumption and being unable to operate efficiently in practical applications.
2. **Small - target loss problem**: In high - level feature maps, small targets are easily submerged by background features, leading to detection failure or a decline in accuracy.
3. **Multi - scale sample problem**: Samples of different scales may lead to the loss of target information after direct resizing, affecting the detection effect.

To solve these problems, the author proposes the following key modules and techniques:

### 1. Lightweight Feature Extraction Attention Module (LFEA)
The LFEA module fully extracts target features through a dual - branch path and uses the ECA module to further strengthen the information interaction between channels. This ensures that the network can effectively extract features while remaining lightweight.

### 2. Low - level Feature Distribution Strategy (LFD)
The LFD strategy gradually distributes low - level features into high - level feature maps to supplement information and reduce the loss of small targets in high - level feature maps. In addition, a Simplified Bilinear Interpolation Attention Module (SBAM) is introduced to promote the fine - grained fusion of low - level features to high - level features.

### 3. Refined Feature Transfer Module (RFT)
The RFT module improves the feature extraction and transfer capabilities through depth - wise separable convolution and ECA layers while remaining lightweight. Compared with direct cross - layer connections, the RFT module can perform more effective feature refinement extraction.

### 4. New training and inference cropping strategies
The traditional method of directly resizing images is abandoned, and a new cropping strategy is adopted. In the training phase, samples are randomly cropped to 256×256 pixels, and in the inference phase, a sliding - window cropping method is used, so as to better adapt to multi - scale samples and improve detection accuracy.

Through these techniques, LR - Net achieves an effective balance between detection accuracy and resource consumption and has won the third place in the "ICPR 2024 Infrared Small - target Detection Challenge under Resource - Constrained Conditions".