MBG-PSENet: A PSENet-Based Model for Dense Text Detection in Customs Declarations

Customs declaration images typically contain dense text, complex table structures, and multi-scale fields, as well as degradation factors such as scan noise, stamp occlusion, and low contrast, making it challenging to locate text regions and separate adjacent text instances. To address the issues of high model complexity, insufficient multi-scale feature fusion, and computational redundancy in the detection heads of the standard PSENet for structured customs declaration text detection, we propose a lightweight dense text detection model, MBG-PSENet. This model retains PSENet’s progressive scale expansion mechanism, adopts MobileNetV3 in place of ResNet-50 as the backbone network to reduce the number of parameters and computational overhead, introduces BiFPN to enhance bidirectional multi-scale information exchange between features at different levels, and designs a lightweight detection head, PSEGhostHead, based on GhostModule to reduce redundant convolutional computations. Experiments were conducted on the self-constructed customs declaration text detection dataset CDTD, which contains 879 customs declaration images, including 702 training images and 177 testing images, as well as the ICDAR2015 public dataset. The results show that MBG-PSENet achieves 99.48% Precision, 98.30% Recall, and 98.89% F-measure on the CDTD dataset, with only 1.04 million parameters, 2.05 G FLOPs, a model size of 4.02 MB, and an inference speed of 44.16 FPS. The experimental results demonstrate that MBG-PSENet achieves a good balance among detection accuracy, model size, and computational efficiency, making it suitable for customs declaration OCR applications in resource-constrained scenarios.