Although deep learning models demonstrate superior performance in interpreting 1D electrocardiogram (ECG) data, a vast number of clinical records are archived as static images, limiting the deployment of state-of-the-art models. This study investigates whether to address this by reconstructing the 1D signal from the image or by applying computer vision models directly to the image. Utilizing the PTB-XL dataset and ECG-Image-Kit, we benchmark a direct visual feature extraction approach (using EfficientNet-B0 and DINOv3) against a signal reconstruction pipeline (U-Net followed by a custom 1D ResNet). The models are evaluated on biological age regression, sex classification, and pathology detection. Results indicate that the 1D ResNet model operating on reconstructed signals consistently outperforms 2D vision-based models across all tasks, despite having significantly fewer parameters. For instance, the 1D model achieved an age regression mean absolute error (MAE) of 9.09 years compared to over 14.30 years for the 2D models. The findings suggest that 1D temporal representations of ECG data are more information-dense for diagnostics, and that targeted signal processing remains a more robust framework than direct image analysis using general-purpose foundation models.