Near-InfraRed (NIR) imaging technology plays a pivotal role in assisted driving and safety surveillance systems, yet its monochromatic nature and deficiency in detail limit its further application. Recent methods aim to recover the corresponding RGB image directly from the NIR image using Convolutional Neural Networks (CNN). However, these methods struggle with accurately recovering both luminance and chrominance information and the inherent deficiencies in NIR image details. In this paper, we propose grayscale-assisted RGB image restoration from NIR images to recover luminance and chrominance information in two stages. We address the complex NIR-to-RGB conversion challenge by decoupling it into two separate stages. First, it converts NIR to grayscale images, focusing on luminance learning. Then, it transforms grayscale to RGB images, concentrating on chrominance information. In addition, we incorporate frequency domain learning to shift the image processing from the spatial domain to the frequency domain, facilitating the restoration of the detailed textures often lost in NIR images. Empirical evaluations of our grayscale-assisted framework and existing state-of-the-art methods demonstrate its superior performance and yield more visually appealing results. Code is accessible at: https://github.com/Yiiclass/RING

Denoising (DN) and demosaicing (DM) are the first crucial stages in the image signal processing pipeline. Recently, researches pay more attention to solve DN and DM in a joint manner, which is an extremely undetermined inverse problem. Existing deep learning methods learn the desired prior on synthetic dataset, which limits the generalization of learned network to the real world data. Moreover, existing methods mainly focus on the raw data property of high green information sampling rate for DM, but occasionally exploit the high intensity and signal-to-noise (SNR) of green channel. In this work, a deep guided attention network (DGAN) is presented for real image joint DN and DM (JDD), which considers both high SNR and high sampling rate of green information for DN and DM, respectively. To ease the training and fully exploit the data property of green channel, we first train DN and DM sub-networks sequentially and then learn them jointly, which can alleviate the error accumulation. Besides, in order to support the real image JDD, we collect paired raw clean RGB and noisy mosaic images to conduct a realistic dataset. The experimental results on real JDD dataset show the presented approach performs better than the state-of-the-art methods, in terms of both quantitative metrics and qualitative visualization.

Visible and infrared image fusion (VIF) aims to combine information from visible and infrared images into a single fused image. Previous VIF methods usually employ a color space transformation to keep the hue and saturation from the original visible image. However, for fast VIF methods, this operation accounts for the majority of the calculation and is the bottleneck preventing faster processing. In this paper, we propose a fast fusion method, FCDFusion, with little color deviation. It preserves color information without color space transformations, by directly operating in RGB color space. It incorporates gamma correction at little extra cost, allowing color and contrast to be rapidly improved. We regard the fusion process as a scaling operation on 3D color vectors, greatly simplifying the calculations. A theoretical analysis and experiments show that our method can achieve satisfactory results in only 7 FLOPs per pixel. Compared to state-of-the-art fast, color-preserving methods using HSV color space, our method provides higher contrast at only half of the computational cost. We further propose a new metric, color deviation, to measure the ability of a VIF method to preserve color. It is specifically designed for VIF tasks with color visible-light images, and overcomes deficiencies of existing VIF metrics used for this purpose. Our code is available at https://github.com/HeasonLee/FCDFusion.

Language-guided fashion image editing is challenging, as fashion image editing is local and requires high precision, while natural language cannot provide precise visual information for guidance. In this paper, we propose LucIE, a novel unsupervised language-guided local image editing method for fashion images. LucIE adopts and modifies recent text-to-image synthesis network, DF-GAN, as its backbone. However, the synthesis backbone often changes the global structure of the input image, making local image editing impractical. To increase structural consistency between input and edited images, we propose Content-Preserving Fusion Module (CPFM). Different from existing fusion modules, CPFM prevents iterative refinement on visual feature maps and accumulates additive modifications on RGB maps. LucIE achieves local image editing explicitly with language-guided image segmentation and mask-guided image blending while only using image and text pairs. Results on the DeepFashion dataset shows that LucIE achieves state-of-the-art results. Compared with previous methods, images generated by LucIE also exhibit fewer artifacts. We provide visualizations and perform ablation studies to validate LucIE and the CPFM. We also demonstrate and analyze limitations of LucIE, to provide a better understanding of LucIE.

As the detection accuracy of salient objects in the image has been improved, the focus of research has gradually shifted to how to improve the accuracy of small object detection. However, existing object detection methods mainly study the general object detection using visible images as the input, and most small object detection methods are designed for visible images, leaving the small object detection in infrared images underexplored. Compared with standard scale objects, infrared small objects lack color information, which makes them more dependent on contextual information. In this paper, an infrared small object detection model is proposed based on the standard YOLOv5 model. The local information around small objects is effectively combined with global information by a Dynamic Contextual Information Extraction Module, which adapts to the subtle morphological changes of infrared small objects dynamically. A Channel-Detail Attention Module is designed to aggregate the channel and detail information of the infrared small objects to improve the accuracy of the regression. Considering the problem of loss of detailed features in the process of network convolution, features with new scales are upsampled and fused with shallow features to capture more detailed information of infrared small objects and avoid feature blending. To demonstrate the effectiveness of the proposed method, experiments are conducted on the public infrared datasets, including ITTD, IRSTD-1k, and NUAA-SIRST. The experimental results show that the proposed method outperforms the compared methods in terms of mAP by 5.1% on the ITTD dataset, and the mAP is also improved by 3.7% compared to that of the baseline method (i.e., YOLOv5) .