Underwater image enhancement refers to the process of improving the quality and visibility of images captured underwater. When an image is taken underwater, it is often affected by various factors such as water turbidity, scattering of light, color distortion, and reduced contrast. These factors result in images with poor visibility, low contrast, and a bluish or greenish tint. Underwater image enhancement techniques aim to mitigate these issues and enhance the visual quality of underwater images.
We focus on the research of underwater image enhancement and investigate the relationship between underwater image enhancement and other underwater vision tasks such as underwater object detection.
Low-light image enhancement refers to the process of improving the quality and visibility of images captured under conditions of insufficient lighting. In low-light situations, the captured images often suffer from issues such as noise, reduced contrast, loss of details, and overall poor visibility. Low-light image enhancement techniques aim to address these challenges and enhance the visual quality of images taken in low-light environments.
Our research covers RGB and RAW data-based low-light image enhancement. We also focus on light correction and image denoising.
Image denoising is a process of reducing or removing noise from digital images. Noise in an image refers to random variations in pixel values that are not part of the original image content. These variations can be caused by various factors such as sensor limitations, transmission interference, or other environmental factors. The goal of image denoising is to enhance the quality of an image by suppressing or eliminating the noise while preserving the important details and structures of the image.
Our main focus is RAW data-based image denoising.
Image dehazing is a process of reducing or removing the atmospheric haze or fog-like effects that degrade the visibility and quality of images. Haze occurs when light interacts with particles and molecules in the atmosphere, causing scattering and absorption. In images captured in hazy conditions, objects farther away appear less distinct and have reduced contrast compared to objects closer to the camera. The goal of image dehazing is to recover the underlying scene information and enhance the visibility of objects by estimating and removing the haze or fog.
Our recent focus of image dehazing is the performance in real world.
Blind face restoration refers to the process of automatically restoring or enhancing the appearance of human faces in images without any prior knowledge or explicit guidance about the specific facial characteristics or the intended restoration outcome. The term "blind" signifies that the restoration algorithm operates without explicit information or annotations about the face in the input image. Blind face restoration algorithms aim to improve the visual quality of facial images by addressing common issues such as noise, blur, low resolution, and other distortions.
We study blind face restoration including image and video-based face restoraiton. We are also doing research on reference-based face restoration.
Video inpainting, also known as video completion or video hole filling, is the process of automatically filling in missing or corrupted regions in a video sequence. It involves reconstructing the missing content in a visually plausible and coherent manner, based on the information available in the surrounding frames or the temporal context of the video.
We focus on the video inpainting and object removal.
Flare removal refers to the process of reducing or eliminating lens flare from images. Lens flare occurs when unwanted light scatters or reflects within the camera lens system, resulting in bright spots, streaks, or haze in the image. Flare can occur due to various factors, such as strong light sources in the frame, direct sunlight hitting the lens, or internal reflections within the lens elements. Flare removal techniques aim to mitigate the negative effects of lens flare and restore the original image quality.
The main foucs of this research direction is nighttime flare removal, particularly in the application on mobile devices.
Exposure correction, also known as exposure adjustment, is the process of modifying the brightness and tonal distribution of an image to achieve a desired overall exposure level. The goal is to enhance the visibility of details, improve the image's dynamic range, and create a visually pleasing result. Exposure correction can be performed in various ways, depending on the specific needs and characteristics of the image.
In addition to the traditional exposure correction, we also explore the extremely over-exposed image correction.
Photo retouching, also known as image retouching, is the process of altering and enhancing photographs to improve their visual quality, correct imperfections, and achieve the desired aesthetic outcome. It involves making various adjustments and modifications to the image, ranging from basic fixes to more advanced editing techniques.
We study aesthetics-driven photo retouching and editing.
Under-Display Camera (UDC) image restoration refers to the process of enhancing the quality and visual appearance of images captured by cameras located underneath the display of electronic devices, such as smartphones or tablets. UDC technology allows for a full-screen display without the need for a visible camera notch or bezel. Since the camera is placed underneath the screen, the image quality can be affected by factors such as reduced light transmission, diffraction, and interference from the display pixels. Therefore, UDC image restoration techniques aim to compensate for these limitations and improve the overall image quality.
We explore the potential of UDC and provide the solutions of UDC image synthetis and restoration.
Image colorization refers to the process of adding color to black and white or grayscale images to create a colorized version. It involves assigning appropriate colors to different elements and areas within the image based on various factors, such as context, knowledge of the subject matter, and artistic interpretation. Image colorization involves a degree of interpretation and subjectivity, as the original colors may not be known or accurately preserved in black and white images. The process often relies on historical knowledge, cultural references, and artistic choices. Colorization can bring new life to old photographs, add visual appeal, and provide a fresh perspective on historical or nostalgic imagery.
We focus on unconditional, prompt-based, hint-point-based, reference-based image and video colorization.
In the context of image restoration, the term "backbone" refers to the fundamental architecture or framework used in deep learning models for image restoration tasks. The backbone network forms the core of the model and is responsible for extracting meaningful features from input images. The choice of the backbone architecture depends on the specific image restoration task and the complexity of the problem. Different architectures have been developed and adapted for image restoration, each with its own strengths and characteristics.
We focus on the design of image restoration backbones, particularly in the frequency domain.