Enhancing Image Quality of Aging Smartphones Using Multi-scale Selective Kernel Feature Fusion Network

Abstract

Smartphones are now the most popular medium for photography. Smartphones can capture better images than the hardware they use thanks to Machine Learning-based computational photography. Even though modern phones have sophisticated image enhancement applications, older and aged phones lag behind for a variety of reasons. Lens quality degradation, which results in washed out, soft-looking images, is one of the most common issues seen in older devices. We propose a method for improving such images by using an attention-based multi-scale residual neural network trained on a synthetic dataset. We chose two smartphones: an old device that captures degraded images and a modern flagship that provides reference enhanced images. Then we used the bloom filter, contrast, and highlight adjustment to make the reference images appear degraded. Later, we trained the model using synthetically degraded images and tested it on a variety of older devices. We achieved a maximum Peak Signal to Noise Ratio (PSNR) score of 74 throughout our experiments. To evaluate the model's images, we used Blind Image Quality Assessment (BIQA) methods such as HyperIQA. Aside from correcting the lens issue, the model achieves comparatively better sharpness, contrast, and color processing. Our proposed method generalizes well and achieves up to 11.45% improvement on novel devices by utilizing a very limited amount of data.