<p>The intelligent banknote recognition system using deep learning can significantly improve the accuracy and efficiency of cash handling, particularly benefiting visually impaired individuals. However, in practical applications, issues like scale and angle variations, complex backgrounds, wear, and lighting changes can cause inaccuracies in recognizing key features of banknotes. In this paper, we propose an adaptive RoI-aware network that can dynamically capture the regions of interest (RoIs) to ensure that key regions, such as serial numbers, watermarks, and specific patterns, are accurately detected. The proposed network consists of two main components. First, we develop a multiscale deformable convolution module that adjusts the receptive fields to capture multi-scale features based on the shape of the RoIs. This module is designed to effectively handle scale variations, perspective distortions, and other challenges in complex banknote images. Second, a non-local attention module is employed to model global context, enabling the network to focus more precisely on key regions of banknote images. This approach minimizes the effects of irrelevant backgrounds and damaged areas on detection accuracy. Extensive experiments conducted on banknotes from various countries to demonstrate that the proposed model possesses strong generalization capabilities to handle variations in banknote designs and subtle differences across different national currencies. The proposed network significantly outperforms state-of-the-art methods in both detection accuracy and robustness, achieving the highest accuracy rates of 92.29%, 97.42%, 99.63%, and 99.64% on the INR, PLN, BTN, and UAE datasets, respectively.</p>