<p>Logit knowledge distillation can be widely used in various distillation scenarios due to its computational efficiency and ability to handle heterogeneous knowledge. However, its performance is generally inferior to feature knowledge distillation. We argue that the key factor restricting logit knowledge distillation methods lies in the fact that the total amount of knowledge they transfer is insufficient. Ideally, the maximum amount of knowledge learned by a student network in the distillation of logit knowledge is contained in the probability distribution of the final output of its teacher network. Thereby, we propose a knowledge distillation method based on multiscale features to augment the total amount of knowledge from the teacher network, by modifying the structure of the output layer of the teacher network to acquire different scales of feature mappings. A progressive scale-by-scale curriculum learning pattern is designed to adequately learn different scales of knowledge from the teacher network and improve the effectiveness of knowledge transfer. In addition, an adaptive weighting method is introduced to enable the student network to dynamically regulate its knowledge acquisition from the teacher network based on their performance discrepancy. Furthermore, we leverage a multiscale loss of interclass relationships to enhance knowledge transfer capacity by modeling the discrepancy of interclass similarity distributions between teacher and student networks across different scales. Extensive experiments conducted on multiple benchmark datasets show the effectiveness of the proposed method. Our code is available at <a href="https://github.com/Yao-Zikang/MSKD">https://github.com/Yao-Zikang/MSKD</a>.</p>