Lightweight Infrared Small Target Detection Method Based on Point-to-point Regression

Infrared small target detection， as the core technology of infrared search and tracking systems， plays a crucial role in military precision guidance， low-altitude economic monitoring， border surveillance， and space situational awareness. Its primary objective is to achieve rapid and accurate localization of small targets with low signal-to-noise ratios and weak contrast under long-distance conditions. In recent years， detection methods based on deep learning have gradually become a research hotspot and have achieved remarkable progress in detection accuracy. Most mainstream approaches adopt mask-to-mask or box-to-box regression training strategies， which can effectively enhance model robustness with the support of large-scale annotated datasets. However， these methods suffer from two major drawbacks. On the one hand， they heavily rely on high-quality， large-scale annotations， which are costly to obtain. On the other hand， in order to further improve accuracy， these methods often increase the depth and width of the network to expand model capacity， leading to massive parameters and complex architectures. To mitigate these issues， some studies have attempted to use weak supervision deep learning as a replacement for traditional mask or bounding-box annotations to reduce annotation costs. However， due to the limited granularity of supervision， the detection accuracy is often significantly constrained， making it difficult to balance practicality and precision.To address the above challenges， this paper proposes a lightweight infrared small target detection algorithm based on point-to-point regression. The goal is to achieve high-accuracy detection while significantly reducing annotation and computational costs. The main contributions of this work are as follows. First， an asymmetric full-dimensional dynamic feature extraction module is designed. This module integrates channel shuffle with omni-dimensional dynamic convolution， which effectively models fine-grained dependencies across channel and spatial dimensions while maintaining network lightweightness. As a result， it enhances the discriminative features of the target and suppresses background interference. Second， a high-resolution cross-feature enhancement module is proposed. This module aligns multi-level features via upsampling， downsampling， and convolution mapping， and builds cross-layer interaction branches to dynamically fuse low-level detail features with high-level semantic features. In this way， it strengthens semantic representation while preserving spatial detail information， making it particularly suitable for weak and small target detection scenarios. Finally， an adaptive point regression detection head is introduced. This head models the distribution of target centers using Gaussian heatmaps， and combines local maxima filtering with a scale-adaptive mechanism to achieve precise prediction and sub-pixel localization. In this process， only single-point annotations are required for training， which substantially reduces the cost of data labeling.Extensive experiments conducted on two public datasets， IRSTD-1K and NUDT-SIRST， demonstrate the effectiveness of the proposed method. With an extremely small number of parameters （only 0.24 M） and low computational complexity （1.53 GFLOPs）， the proposed algorithm achieves outstanding performance： F1 score of 91.79% on IRSTD-1K and 97.44% on NUDT-SIRST， which are significantly superior to existing state-of-the-art methods. In addition， the model reaches an inference speed of 36 FPS on a single GPU， fully demonstrating its potential for real-time infrared small target detection in practical applications.