Surgical robot visual perception enhancement method based on generative adversarial network (GAN)

The accuracy of vision perception of surgical robot directly influences the safety and success rate of operation. A new visual perception enhancement method for surgical robot is proposed based on image blur and difficult to extract key areas based on Improved Generation Adversarial Network (GAN). This algorithm uses multiscale feature fusion generator to efficiently extract features from different levels of surgical images. By using dual discriminator, global structure and local detail quality can be guaranteed. Experiments were performed on open surgery data sets and self-built data sets with PSNR, SSIM and mAP as evaluation indexes. Compared with traditional histogram equalization algorithm, PSNR increased by 19.9% from 27.1 to 32.5; SSIM increased 14.1% from 0.78 to 0.89; and mAP increased by 13.6% from 0.81 to 0.92. In comparison to CycleGAN, PSNR increased 15.2%, SSIM 8.6% and 9.5% higher mAP. As far as subjective visual effect is concerned, the image processed by this algorithm has clear edges and rich texture details, so that it can effectively enhance visual perception ability of surgical robots. However, this algorithm suffers from real time performance limitations, and average processing times for each high resolution surgical image are 0.8s.