Deep Learning for Image Restoration and Robotic Vision.

다국어 초록 (Multilingual Abstract)

Traditional model-based approach requires the formulation of mathematical model, and the model often has limited performance. The quality of an image may degrade due to a variety of reasons: It could be the context of scene is affected by weather conditions such as haze, rain, and snow; It's also possible that there is some noise generated during image processing/transmission (e.g., artifacts generated during compression.). The goal of image restoration is to restore the image back to desirable quality both subjectively and objectively. Agricultural robotics is gaining interest these days since most agricultural works are lengthy and repetitive. Computer vision is crucial to robots especially the autonomous ones. However, it is challenging to have a precise mathematical model to describe the aforementioned problems. Compared with traditional approach, learning-based approach has an edge since it does not require any model to describe the problem. Moreover, learning-based approach now has the best-in-class performance on most of the vision problems such as image dehazing, super-resolution, and image recognition. In this dissertation, we address the problem of image restoration and robotic vision with deep learning. These two problems are highly related with each other from a unique network architecture perspective: It is essential to select appropriate networks when dealing with different problems. Specifically, we solve the problems of single image dehazing, High Efficiency Video Coding (HEVC) loop filtering and super-resolution, and computer vision for an autonomous robot. Our technical contributions are threefold: First, we propose to reformulate haze as a signal-dependent noise which allows us to uncover it by learning a structural residual. Based on our novel reformulation, we solve dehazing with recursive deep residual network and generative adversarial network which emphasizes on objective and perceptual quality, respectively. Second, we replace traditional filters in HEVC with a Convolutional Neural Network (CNN) filter. We show that our CNN filter could achieve 7% BD-rate saving when compared with traditional filters such as bilateral and deblocking filter. We also propose to incorporate a multi-scale CNN super-resolution module into HEVC. Such post-processing module could improve visual quality under extremely low bandwidth. Third, a transfer learning technique is implemented to support vision and autonomous decision making of a precision pollination robot. Good experimental results are reported with real-world data.