An unsupervised transfer method for composited image to real image under dark light conditions

Most of the existing hand data collection procedures rely on complex and bulky image acquisition systems, in fact, the data directly acquired by the acquisition systems do not have a realistic context, and the neural network models trained from these data perform not well in real-world situations, especially when people are wearing gloves in scenes with changing illumination. However, to improve the consistency of the image the traditional image fusion methods sacrifice the authenticity of the background and the supervised methods are limited by the data acquisition system which cannot generate high quality fusion image. Therefore, based on an unsupervised generative adversarial network, this paper introduced a generator based on a global attention module and proposed a histogram similarity-based image selection module to filter the input images. Our goal is to decrease the difficulty of the image migration task for the network. The model had been trained using a self-built dataset consisting of composited and real data, including images of hands wearing IMU data gloves and photographs taken in real life with similar backgrounds. Extensive experiments on image segmentation tasks demonstrated the effectiveness of the proposed model, which obtained a precision of 0.88 and a recall of 0.86 with a mean NIMA value of 4.27 compared to other state-of-the-art methods such as UEGAN and DoveNet.