Advanced Text Detection of Container Numbers via Dual-Branch Adaptive Multi-Scale Network

Text detection of container numbers is essential in logistics for the efficient tracking and management of containers. However, the wear and damage that container numbers often endure in harsh environments make detection particularly challenging. We propose a Dual-Branch Adaptive Multi-Scale Network (DAMNet) to address these issues. The DAMNet integrates Efficient Multi-Scale Attention (EMA) into the Differentiable Binarization (DBNet) model to address the challenges of improving accuracy and efficiency in the current text detection networks. We also incorporate an auxiliary information branch that processes images using the Otsu thresholding method to extract local features. This additional branch enhances the model’s ability to learn the features of container number text. Furthermore, we improve the feature pyramid structure by introducing Adaptive Spatial Feature Fusion (ASFF). This method retains rich detailed information while capturing global context, improving the model’s robustness. We also propose a dual-loss strategy. It employs BCE loss for classification tasks and Dice loss to tackle imbalanced positive and negative samples. Both losses are combined via weighting. The experimental results demonstrate that the improved DBNet model achieves a precision of 89.7%, a recall of 80.1%, and an F-score of 83.5% on the ICDAR2015 dataset. The recall increased by 3.9% compared to the original model.