Color Restoration Method for Endoscope Image Using Multiscale Discriminator Based Model Compression Strategy

Color restoration of endoscopic images is an urgent clinical need during photodynamic surgery. In recent years, deep learning methods achieved notable results in the fields of image processing. The model compression algorithm and hardware performance enhancement improved the model inference speed. It is possible to apply deep learning methods to the task of endoscopic image color restoration during surgery. However, experiments show that model compression can lead to image deterioration. To solve this issue, we propose a fast color restoration method for endoscopic images, which use multiscale discriminator based on model compression. Initially, we train a CycleGAN teacher network with multiscale discriminator. Then, we obtain the student model through knowledge distillation and neural architecture search. We use the trained teacher discriminator to guide the student model and add feature matching loss to stabilize the training process. Experiments show that our method ameliorates the performance of compressed models.