Workshop on Metrification and Optimization of Input Image Quality in Deep Networks (MOI2QDN)

摘要

Recent years have seen significant advances in image processing and computer vision applications based on Deep Neural Networks (DNNs). This is a critical technology for a number of real-time applications including autonomous vehicles, smart cities, and industrial computer vision. Often deep neural networks for such applications are trained and validated based on the assumption that the images are artefact-free. However, in most real-time embedded system applications the images input to the networks, in addition to any variations of external conditions, have artefacts introduced by the imaging process, such as the Image Signal Processing (ISP) pipelines. Data augmentation methods are exploited to expand the subset of trained images to include images with different levels of distortions. Such data augmentation often may result to improved performance of the network for artefact-specific images but also degrades the performance for artefact-free images.