Digital ink Chinese character recognition (DICCR) systems have predominantly been developed using datasets composed of native language writers. However, the handwriting of foreign students, who possess distinct writing habits and often make errors or deviations from standard forms, poses a unique challenge to recognition systems. To address this issue, a robust and adaptable approach is proposed, utilizing a residual network augmented with multi-scale dilated convolutions. The proposed architecture incorporates convolutional kernels of varying scales, which facilitate the extraction of contextual information from different receptive fields. Additionally, the use of dilated convolutions with varying dilation rates allows the model to capture long-range dependencies and short-range features concurrently. This strategy mitigates the gridding effect commonly associated with dilated convolutions, thereby enhancing feature extraction. Experiments conducted on a dataset of digital ink Chinese characters (DICCs) written by foreign students demonstrate the efficacy of the proposed method in improving recognition accuracy. The results indicate that the network is capable of more effectively handling the non-standard writing styles often encountered in such datasets. This approach offers significant potential for the error extraction and automatic evaluation of Chinese character writing, especially in the context of non-native learners.