To address the challenges posed by an aging population and to enable bathing robots to apply appropriate cleaning modes for different skin regions, the research was conducted on multi-part skin detection and its application in these robots. Building on previous studies, four representative object detection algorithms were selected, the original dataset was expanded, and multi-part skin detection was performed using transfer learning. A comprehensive evaluation metric was developed to assess algorithm performance, and the best-performing model was deployed and tested on Tesla T4 and TX2 platforms, with practical applications in bathing robots. The results showed that addressing dataset class imbalance lead to an average accuracy improvement of 18%. YOLOv5s has achieved an optimal balance between accuracy and model size, facilitating real-time detection on both Tesla T4 and TX2, and effectively identifying different skin regions in a humid environment. By integrating a visual sensor with TX2, 3D pose modeling and joint experiments were conducted, achieving a 92% success rate in controlling the robot to reach the back area. Utilizing point cloud data as supervision improve the success rate to 100%. Addressing class imbalance significantly enhances the accuracy of multi-part skin detection, and YOLOv5s exhibites excellent performance in balancing accuracy and model size, effectively meeting the multi-part skin detection requirements of bathing robots.