Traditional Knowledge Distillation (KD) methods efficiently transfer knowledge to a lightweight student model with the help of a well-trained teacher model. Further, Deep Mutual Learning (DML) introduces a collaborative training paradigm wherein dual peer student networks engage in reciprocal knowledge exchange, demonstrating that model performance can be enhanced without reliance on a strong teacher network, while simultaneously reducing architectural and computational complexity. In this paper, we introduce the Self Mutual Learning(SML) framework for scene text detection aimed at further developing DML. First, SML further simplifies the training process. Only build different detection heads, and promote mutual learning between different detection heads to improve the model performance. Second, to increase the diversity of detection heads, we train a pruner to make the input features of the additional detection head sparse. With the help of the pruner, the additional detection head is forced to use fewer features to predict the result, which benefits the model to capture key information. At the same time, due to the different input features, detection heads can provide richer knowledge to each other. Extensive experiments demonstrate that our method consistently improves the F-measure of the model and outperforms the other mainstream distillation methods.