Research on the Application of Deep Learning-Based Product Spray Code Text Recognition Algorithm

In the process of product manufacturing, the characters printed on the packaging may be incomplete due to reasons such as rapid mechanical vibrations, low ink quality, and poor operating environments. Therefore, manufacturers often perform defect detection of printed codes during the product quality inspection phase. To achieve efficient recognition of printed codes in complex environments, this paper adopts a sequence-based CRNN text recognition algorithm and designs a lightweight network structure. First, MobileNetV3_small is used as the feature extraction network, optimizing the sampling stride to better preserve the character features in the image. Then, a bidirectional Long Short-Term Memory network (Bi-LSTM) is used to predict the text sequence, combined with the CTC loss function to achieve end-to-end variable-length text sequence recognition. To verify the performance of the improved algorithm, experiments were conducted on two commonly used text recognition datasets,IIIT5K-Words and ICDAR2013. The experimental results show that the combination of CRNN and MobileNetV3_small not only effectively improves the speed of text recognition but also ensures the accuracy of printed code recognition, thereby enhancing the effectiveness of quality inspection for printed codes.