Transformer-based Image Classification Model with Multi-scale Dynamic Window

In recent years, Transformer-based vision models have achieved remarkable progress in image classification tasks. However, conventional Vision Transformers (ViTs) encounter two major challenges when dealing with complex visual scenes: the fixed-size patch partitioning restricts the model’s capacity to capture fine-grained local information, and the self-attention mechanism incurs high computational complexity. To address these issues, this paper explores a dynamic window-based multiscale transformer model (DWMTransNet) for image classification. By employing receptive fields of varying sizes across different attention heads, the proposed model adaptively captures multiscale features. Additionally, a dynamic window mechanism is introduced, enabling the model to flexibly adjust its feature extraction strategy according to the image content. Experiments conducted on the CIFAR-10 dataset demonstrate that, compared to the standard ViT, the DWMTransNet achieves an approximate 13% improvement in classification accuracy (reaching 77.34%), while simultaneously reducing computational cost by around 22%. The experimental results validate the effectiveness of the proposed method in enhancing both classification performance and computational efficiency.