A Network of Generating High Quality Hemoglobin Images

In recent years, image-based facial skin pigment extraction has made great progress, and the extraction methods are mainly divided into traditional methods and deep learning methods. Due to the early introduction of traditional methods, there are some defects, such as low efficiency, some black pixels will go white, etc. Therefore, this paper mainly proposes a deep learning method based on Generative Adversarial Network (GAN) to achieve hemoglobin image generation. Our goal is similar to image translation. Currently, the mainstream GAN networks for image translation include pix2pix, pix2pixHD, cycleGAN, etc. This paper proposes a network GHNet based on UNet to generate higher quality images. In the data set part, we built the paired data set of polarized light-hemoglobin image by ourselves. In the network part, we reduced the number of encoders and decoders in the traditional UNet network, which made the network more lightweight. Based on this, we define a neural network module and integrate the attention mechanism SE module to enhance the ability to capture the input image features. The experimental results show that, compared with ASAPNet, DGC and CLD networks, the proposed network training speed is the same as that of the baseline network, and the PSNR index and SSIM index of the generated images are improved.