Human-guided robot joints self-learning and transfer learning analytics

Style transfer techniques have seen wide adoption in recent years, with the CUT and CycleGAN network structures standing out for their superior accuracy and robustness compared to other methods. However, conventional style transfer methods rely on the structural characteristics of the content image to guide the transfer, resulting in a style image that mimics the original image structure. To address this limitation, this study proposes a novel approach that utilizes mo- tion information of the human body as transfer information for style transfer networks. Specifically, the goal is to adapt the style of motion between a hu- man hand and a robotic arm, thereby generating a virtual robotic arm image that reflects the same motion style as the human hand. The proposed method combines depth curve estimation-based image processing techniques with style transfer to enable stable stylized operations on content images under varying lighting conditions, thereby enhancing the robustness of the overall approach. To evaluate the proposed method, a suite of quantitative metrics, including PSNR, SSIM, MAE, L1, FID, and LPIPS, are employed to analyze the test results of the model, both before and after optimization. Future research could potentially explore incorporating object recognition techniques to extract target objects for the model. This study’s approach holds positive implications for the field of robotics, especially in the area of robot training. Keywords: Style transfer, depth curve processing, robotics, image processing, computer vision, generative adversarial networks, contrastive unpaired translation, transfer learning, human perception.Master of Science (Computer Control and Automation