Physically inspired geometry-to-photometry rendering and lightweight convolutional neural network for real-time weld defect detection

Abstract

Welding quality inspection of battery top covers is critical to lithium-ion battery safety and reliability. The height deviation between weld defects and normal areas is extremely small; on specular, weakly textured surfaces it produces no stable brightness contrast, so two-dimensional (2D) images carry almost no usable defect signal. Three-dimensional (3D) sensing captures such deviations but incurs computational costs prohibitive for production lines. Raw depth maps, as scalar grayscale fields, cause 2D convolutional networks to overlook tiny-defect features. Addressing all three limitations simultaneously remains a key challenge for artificial intelligence (AI) in industrial inspection. To address this challenge, we propose a geometry-to-photometry defect detection framework that converts 3D weld geometry into discriminative 2D photometric representations via physically inspired rendering. Depth data are first mapped to pseudo-color images, and surface normals then modulate lighting under a Phong-based model to transform subtle geometric discontinuities into high-contrast photometric cues. Building on this strategy, we develop YOLO-WaveletTiny (YOLO-WT), a lightweight You Only Look Once (YOLO) detector integrating wavelet depthwise separable convolution and inverse wavelet upsampling for efficient feature extraction and reconstruction. Experiments on a real industrial dataset with six defect categories show the rendered representation improves the baseline by 5.9 percentage points in mean average precision at 50% intersection over union (mAP@50), and the full framework achieves 93.8% mAP@50 with only 2.47 million parameters and 5.7 GFLOPs, outperforming mainstream detectors. These results demonstrate the framework as an effective, practical AI solution for low-cost, high-precision weld defect inspection.