Online signature verification captures dynamic information during the signing process to identify and authenticate signatures. The extraction of effective signature features has been a central focus in this area of research. In this paper, a novel online signature verification model is proposed, which combines multi-scale convolution with dynamic time warping (DTW). The model extracts multi-scale features from signatures using a multi-scale convolutional network (MSCN), and adaptively weights different feature channels using a channel attention mechanism. This enables the model to extract more effective features from signatures of varying scales, thus enhancing the model’s representational capacity. Furthermore, soft-DTW is used as a distance metric, and compared to DTW, soft-DTW can address the non-differentiability issue of DTW and enable end-to-end training. Experimental results demonstrate that the proposed model significantly outperforms previous methods on the DeepSignDB dataset, achieving state-of-the-art results.