To address the issue of local feature misalignment caused by modality differences between photo and sketch images in sketch face recognition, a novel method based on local feature reconstruction was proposed. This approach leveraged face labels mapped to learnable text tokens to reduce the distance between matched image pairs, thereby enhancing the model recognition performance. A parameter-free local feature alignment module was introduced, which reconstructed the facial region of one modality using the local features of the other modality, without relying on external references, to achieve more precise cross-modal local feature alignment. In order to fully utilize the cross-modal description capability of visual language pre-training, a two-stage training strategy was proposed. In the first stage, the image and text encoders were frozen, and only the text tokens were optimized. In the second stage, the text tokens specific to face labels and their encoders were kept fixed, and only the image encoder was fine-tuned. Experimental results show that the proposed method outperforms current mainstream approaches on three public sketch face datasets.