Dim infrared image enhancement based on convolutional neural network

Long-range infrared images are always suffering from dim targets and background clutters. To improve the contrast between target and background, we propose a novel infrared image enhancement approach by highlighting target and suppressing background clutters. Predicting the target and background plays a key role in improving the contrast of dim infrared images that targets are embedded by background clutters. Taking full advantage of machine learning on prediction, we design the convolutional neural network (CNN) architecture in our study. To overcome the lack of large training data, the handwritten images in MNIST dataset are employed to simulate the properties of long-rang infrared images including dim targets, background clutters and low contrast. The target and background sub-images are predicted from the original dim infrared image based on the filters in the first layer of the trained CNN. Finally, the dim infrared image is enhanced by amplifying the targets and subtracting background clutters. The results of subjective and quantitative tests prove the performance of the proposed algorithm in contrast improvement.