Doctoral Seminar: CNN-based image / video enhancement and compression

Speaker: Kai Cui

Image/video enhancement is a critical component of image acquisition/processing pipelines. Motivated by the special characteristics of the inter-channel correlation and general characteristics of distorted color images, a 3-stage convolutional neural network structure is proposed to solve color image enhancement tasks. This talk first gives an overview of the 3-stage CNN structure and its variants, and then analyze the proposed structure with three specific color image enhancement tasks, including color image demosaicking, color compression artifacts reduction, and real-world color image denoising. For each of them, an in-depth analysis is performed, which brings a more comprehensive understanding of the proposed structure and provides insights for future work.