Lightweight Multi-Scale Semantic Segmentation of TOFD Weld-Defect Images

Time-of-flight diffraction (TOFD) images contain small defect indications and strong geometric signals that complicate automated interpretation. This study evaluates a lightweight segmentation configuration using seven observed classes: background, four defect-indication categories, head wave, and bottom signal. Eight models were trained on the same fixed partition with three seeds (42, 43 and 44). Enhanced TOFD-Net achieved a mean intersection over union (mIoU) of 0.8854 ± 0.0039 and a defect-union IoU of 0.8459 ± 0.0040 with 3.84 million parameters and 115.11 million multiply–accumulate operations. DeepLabV3-ResNet50 attained the highest mIoU, 0.8909 ± 0.0034, with 39.63 million parameters. A separate five-fold analysis describes image-level variability and mask-component errors. The findings support an accuracy–size trade-off on the evaluated dataset. Image-index grouping does not verify specimen- or scan-level independence, and evaluation informed sampler development. Extended internal-indication (LBIN) annotations occur only in simulated images. Pixel agreement and mask-component statistics do not establish physical flaw detection or sizing. CPU and laptop-GPU measurements cover model forward passes; edge deployment and end-to-end inspection performance were not assessed.