Punctate Highlight Removal for Microscopic Images Based on Pix2pixHD

Punctate Highlight Removal for Microscopic Images Based on Pix2pixHD

• Designed a framework based on pix2pixHD to remove punctate highlights in microscopic images.
• Proposed a microscopic image dataset consisting of 7500 microscopic images in the size of 256×256 with punctate highlights and 7500 corresponding images without highlights.
• We combine the traditional method of highlight removal with the deep learning model pix2pixHD. The traditional method is employed to generate a binary mask indicating the approximate highlight area as the priori information. Then the mask is concatenated with the 3-channel highlighted microscopic image as the input of the generator.
• Our model is trained by calculating a weighted GAN loss to achieve a better performance of highlight removal. Specifically, we simply increase the weight of loss of highlighted pixels according to the binary mask.