Detailed radiation dose fields are essential for ensuring nuclear energy safety and protecting personnel inside and outside the station. Traditional forward modeling methods depend on accurate descriptions of radiation sources and extensive prior knowledge, which are often unavailable in complex or dynamic radiation scenarios, such as after nuclear accidents. Radiation field re-construction methods use monitoring data to infer source information and reconstruct field distributions, reducing reliance on prior knowledge and better adapting to complex environments. Most current approaches rely on 2D interpolation algorithms, including Kriging and radial basis functions. While these methods are fast and effective, they require high-quality sample data and perform poorly in complex scenarios. Neural network-based methods, particularly convolutional neural networks (CNNs), have gained attention for their superior data fitting and predictive capabilities. However, existing methods often rely on simple network structures, limiting their ability to handle complex radiation fields. To address this, this study proposes an improved CNN framework for reconstructing complex 2D radiation fields. Results demonstrate a 1000-fold interpolation reconstruction with over 90% of test samples achieving an average relative error below 10%, and most regions exhibiting measurement point errors below 5%. This approach provides a robust solution for high-precision reconstruction in challenging scenarios.