BRING LIGHT TO THE NIGHT: CLASSIFYING THERMAL IMAGE VIA CONVOLUTIONAL NEURAL NETWORK BASED ON VISIBLE DOMAIN TRANSFORMATION

Most existing vision systems target at processing images captured during the day time. However, it is also essential to enable cameras to see the scenes during the night, such as in outdoor places where no light exists and power outage in indoor environments. We capture thermal images to observe objects in the dark environment. Based on the captured thermal images, we develop a convolutional neural network to classify the images As thermal images require to invest a substantial amount of time to create clear images, we also rely on color images to enrich the training samples and apply transfer learning to refine the CNN classification models. The visible source domain network is learned together with a decoding network to enforce the source domain learning outcome resembling the target thermal domain properties.