The recognition accuracy of dot matrix characters is low due to error segmentation, this paper proposes a dot matrix character detection method based on convolutional neural network (CNNs) recognition feedback. Firstly, multi-scale windows are used to acquire multiple candidate regions and CNNs are established to identify them. The voting mechanism is used to make comprehensive decisions on multiple recognition results, and then the lattice characters are reversed according to the decision result and the character segmentation is completed. Finally, a sliding flip window is proposed to segment and identify all characters. The experimental results show that the proposed method outperforms the traditional character recognition method in the segmentation accuracy and recognition rate of dot matrix characters, reaching 97. 53% and 97. 50% respectively.