A fast filter enhancement method for the infrared image

Abstract

Recent advances in image enhancement explored the power of convolutional neural network (CNN) to achieve a better performance. Despite the great success of CNN-based methods, it is not easy to apply these methods to edge devices (such as FPGA, ASIC) due to the requirement of heavy computation, and the CNN-based methods heavily rely on the training datasets. In this paper, we propose a simple method for image enhancement, without any training steps. We tackle a fundamental yet challenging problem to improve the quality of the infrared images. This type of low light is very common during the infrared photo taking. We found existing methods, based on local or global information, cannot improve the quality of the infrared images, which have been designed for RGB images. We propose a simple yet effective filter via two common parts, named structure recovery and noise removal. It directly establishes correspondence between accuracy and speed for the further applications. Extensive experimental results show that the proposed method achieves a better trade-off against the other methods in terms of performance and model complexity. Moreover, our method achieves 65fps on the edge device.