# CTVNet **Repository Path**: weilizhang/CTVNet ## Basic Information - **Project Name**: CTVNet - **Description**: No description available - **Primary Language**: Unknown - **License**: Not specified - **Default Branch**: main - **Homepage**: None - **GVP Project**: No ## Statistics - **Stars**: 0 - **Forks**: 0 - **Created**: 2026-04-24 - **Last Updated**: 2026-04-24 ## Categories & Tags **Categories**: Uncategorized **Tags**: None ## README

CTVNet: Gradient Prior-Guided Deep Unfolding Network for Infrared Small Target Detection

Pei Liu¹ Li Pang¹ Jiangjun Peng² Yisi Luo¹ Junmin Liu¹ Xiangyong Cao^✉1

¹Xi'an Jiaotong University ²Northwestern Polytechnical University

## Abstract > *For infrared small target detection tasks, deep unfolding techniques have demonstrated effectiveness and practical value. However, existing methods generally emphasize the low-rankness of background and the sparsity of targets within the robust principal component analysis (RPCA) framework, which may overlook the intrinsic gradient prior information existed in background. To address the challenges of complex background estimation and accurate small target detection, we propose a gradient prior-guided deep unfolding network, termed the correlated total variation network (CTVNet). First, we introduce a correlated total variation regularization to simultaneously characterize the low-rankness and local smoothness of the background, and transform it into the estimation of gradient maps. Subsequently, we employ a multi-scale feature fusion network to thoroughly extract gradient priors, replacing the complex and limited analytical computation of gradient correlations. Finally, we unfold the designed iterative algorithm using alternating direction method of multipliers (ADMM) into a learnable network, where each module corresponds to a specific operator within the iterative process, and all parameters are learnable. By training the network end-to-end, the learnable modules can be automatically optimized to better separate the background and the target. Extensive experimental results demonstrate that our proposed method achieves competitive performance compared to several state-of-the-art algorithms while exhibiting superior performance and generalization capabilities on both in-distribution and out-of-distribution data.* ## Framework ![CTVNet_model](https://github.com/user-attachments/assets/275681db-a037-4b70-955d-b4eaf349f004) Fig.1 Framework of CTVNet. (a) The iterative process of the proposed deep unfolding network for infrared small target detection. (b) The design of a single stage in our network. Each stage consists of a background module, gradient prior module, and sparse module to accomplish the update of background image $\mathcal{B}$, gradient map $\mathcal{G}$, and target image $\mathcal{T}$, respectively. Furthermore, the auxiliary variables $\varLambda _k, k\in \varGamma $, $\varUpsilon$, and $\mu$ are updated by auxiliary module. ## Details ![grad](https://github.com/user-attachments/assets/f2c58dcd-92f6-422b-9199-1e9e40be5c24) Fig.2 The heatmaps of variable $T^{\left( s \right)}$ and its corresponding 3D maps during the iterative process of CTVNet when $ S = 3 $, selected from typical scenes in the MSISTD and NUDT-SIRST datasets. | ![grad](https://github.com/user-attachments/assets/a633a970-dbab-4433-81e8-a3206d9eed44) | ![alg](https://github.com/user-attachments/assets/40a7b678-1056-49aa-b352-5797407c8037) | |------------------------|------------------------| | Fig.3 Infrared small target image with gradient maps along the X and Y directions and their corresponding histograms. | Fig.4 Deep unfolding algorithm of our proposed CTVNet. | ## Datasets We evaluate our approach on five publicly available datasets: - [MSISTD](https://github.com/Crescent-Ao/MSISTD) - [NUDT-SIRST](https://github.com/YeRen123455) - [SIRST-v1](https://github.com/YimianDai/sirst) - [IRSTD-1K](https://github.com/RuiZhang97/ISNet) - [ATR](http://www.sciencedb.cn/dataSet/handle/902) In general, all images are rescaled to a standardized resolution of $256 \times 256$ pixels for consistency. Specifically, for the IRSTD-1K dataset, directly resizing $512 \times 512$ images to $256 \times 256$ causes excessive target shrinkage, so a random $256 \times 256$ crop is applied to the target region. Additionally, we have strictly partitioned the training, validation, and test sets to better assess the model's generalization, with the dataset division details shown below:. **Table**: The detailed division of the datasets into training, validation, and test sets adopted in the experiment. | **Methods** | **Image Type** | **Total** | **Training** | **Validation** | **Test** | |------------------------|:----------------:|:-----------:|:--------------:|:----------------:|:----------:| | MSISTD [[Wang et al.](https://github.com/Crescent-Ao/MSISTD)] | synthetic | 1072 | 856 | 108 | 108 | | NUDT-SIRST [[Li et al.](https://github.com/YeRen123455)] | synthetic | 1327 | 663 | 332 | 332 | | SIRST-v1 [[Dai et al.](https://github.com/YimianDai/sirst)] | real | 427 | 256 | 85 | 86 | | IRSTD-1K [[Zhang et al.](https://github.com/RuiZhang97/ISNet)] | real | 1001 | 800 | 100 | 101 | | ATR [[Hui et al.](http://www.sciencedb.cn/dataSet/handle/902)] | real | 805 | 619 | 70 | 116 | ## Run the code ``` python train_main.py ```