Manual inspection of high-voltage cable insulation is inefficient, and traditional algorithms falter at weak boundaries. To address this, we propose a portable measurement system and a two-stage “coarse localization-fine reconstruction” algorithm. First, a novel scale-aware attention U-Net (SA-UNet) performs precise segmentation. It incorporates a multi-scale attention fusion encoder (MSAF), a refined multi-kernel pooling (RMP) module, and a Skip attention feature fusion (SAFF) module to enhance perception of complex details and weak boundaries, effectively mitigating issues from class imbalance. Subsequently, a Pix2Pix generative adversarial network (GAN) reconstructs low-contrast regions, sharpening interlayer boundaries by learning a mapping from blurred to clear images. Finally, insulation thickness is automatically computed using the ray intersection method. Validated on a dataset of 3 300 cross-sectional images, the proposed SA-UNet achieves a 99.36% intersection over union(IoU), outperforming models like U-Net and DeepLabv3+. The GAN reconstruction achieves high structural fidelity (SSIM > 0.98), enabling a final thickness measurement with a mean absolute error (MAE) of only 0.01 mm. This work presents a robust, automated solution for high-precision measurement of high-voltage cable insulation parameters.