Abstract: As China's offshore oil and gas resource development shifts from shallow to deep waters, the number of offshore platforms increases. The main structures of these platforms serve long-term in marine environments. They are affected by seawater corrosion and external forces, leading to surface cracks in the support structures. Thus, it is particularly important to evaluate and assess the surface defects of structures. Current methods for sizing are limited to horizontal cracks and struggle with angled cracks. Based on ACFM theory, this paper used theoretical analysis and simulation to study the relationship between crack angles/sizes and magnetic field images. Experiments with different crack angles and sizes validated the simulation results. A convolutional neural network-based algorithm was established for high-precision quantification of crack angles, lengths, and depths. The results showed that the relative error for angle quantification was 2.27%, for straight crack length and depth, the relative error was 0.73% and 2.73%, and for angled crack length and depth, the relative error was 8.07% and 9.56%, which were within the allowable error range.