Learning Enriched Features for Fast Image Enhancement

In the realm of image restoration, the objective is to recover high-quality image content from degraded input images. This pursuit finds applications in diverse fields such as computational photography, surveillance, autonomous vehicles, and remote sensing. Recent years have witnessed notable progress in image restoration, primarily driven by convolutional neural networks (CNNs). Existing CNN based methods are commonly designed to operate either on full-resolution images, preserving spatial details but lacking precise contextual encoding, or on progressively lower-resolution representations, providing semantically reliable outputs with compromised spatial accuracy. This paper introduces a novel architecture with a comprehensive aim of maintaining spatially-precise high-resolution representations throughout the network, while simultaneously incorporating complementary contextual information from low-resolution representations. The central element of our approach is a multi-scale residual block featuring key components: (a) parallel multi-resolution convolution streams for extracting multi-scale features, (b) information exchange across these multi-resolution streams, (c) a non local attention mechanism to capture contextual information, and (d) attention based multi-scale feature aggregation. Our approach learns an enriched set of features by combining contextual information from multiple scales while preserving high-resolution spatial details.