In metal forming and remanufacturing applications, the assessment of residual plastic deformation is a key factor for process robustness, structural integrity, and reuse-oriented decision making. However, conventional full-field measurement techniques such as Digital Image Correlation (DIC) require an undeformed reference configuration, which is often unavailable for end-of-life or pre-deformed components, making these techniques not applicable. To overcome this limitation, this work applies a reference-free, data-driven approach based on Machine Learning to retrieve the strain field over a sample surface directly from a single image of the deformed configuration. The only requirement is that the investigated surface provides sufficient texture information for the learning model, since overly uniform or highly irregular patterns prevent reliable discrimination between deformed and undeformed regions. The method exploits a convolutional neural network derived from the VGG-16 architecture, trained to infer the in-plane principal strains directly from images of deformed surface patterns. The proposed methodology is applied to pre-deformed metallic components relevant to forming and remanufacturing scenarios, including steel and aluminum alloys subjected to different deformation histories. Different surface patterns are investigated, namely material microstructures observed by optical microscopy, surface roughness, and pre-printed speckle patterns similar to those used in DIC. Validation is first performed on synthetic data using an image simulator capable of reproducing realistic images of deformed samples, including the main sources of uncertainty such as noise, illumination variations, and out-of-plane motion. Finally, the approach is applied to real deformed sheet metal components made of aluminum alloys. The method is shown to provide a reliable estimation of the deformation level of the material. A discussion on accuracy and uncertainty sources affecting the proposed approach is also provided. The results highlight the potential of AI-based, reference-free strain estimation as a complementary tool for uncertainty-aware inspection, classification, and decision support in metal forming and remanufacturing processes.