PD-UNet: a pixel-level detection convolutional neural network for key point detection of car body styling surfaces

The conceptual design phase is the most time-intensive stage in car body structure design. To accelerate the design process, many companies shift the design task to selecting from a body-in-white design library, necessitating the extraction of key information from conceptual design sketches. This paper proposes a method for identifying key points from conceptual design sketches, establishing a link between the sketch and the body-in-white feature framework. To solve the problem of identifying feature points in car body styling surfaces, this study introduces a pixel-level detection method for recognizing key points of car body styling surfaces, known as PD-UNet. Integrated with the grayscale conversion and image downscaling method, the effective features of the image are enhanced, and data redundancy is reduced; then a new U-Net structure is proposed to achieve high-accuracy key points detection. Typically, the heatmap is used as the loss function to increase the model’s robustness and stability, thereby improving training efficiency. The trained PD-UNet model is compared to traditional key points detection algorithms. The results indicate that the model's prediction accuracy can reach 93.38%, satisfying the accuracy and stability requirements of detection tasks. This method effectively achieves key points detection in conceptual sketches of car body styling structures.