Container-Number Detection and Recognition in Land-Port Yard Scenes Based on an Improved YOLO11 Detector and a Lightweight CRNN Recognizer

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Automatic container-number recognition in land-port yards is challenged by clutter, illumination changes, scale variation, surface wear, and character distortion. This paper proposes a two-stage detection–cropping–recognition framework comprising an improved YOLO11 detector and a lightweight convolutional recurrent neural network (CRNN). The detector integrates C3k2 with channel-split multiscale aggregation (C3k2-CSMA), a hierarchical efficient feature fusion pyramid network (HEFFPN), selective boundary aggregation (SBA), and a shared enhanced detection head (SED-Head) with detail-enhanced convolution (DEConv) to improve small-target representation, feature fusion, and boundary localization. The recognizer combines MobileNetV3 with bidirectional long short-term memory (Bi-LSTM) and is trained using connectionist temporal classification (CTC) loss and L2 regularization. The detector achieved 97.3% mean average precision at an intersection-over-union threshold of 0.5 (mAP@0.5), 97.1% recall, 55 frames per second (FPS), and a model size of 12.2 MB. The recognizer achieved 97.6% exact-match accuracy, a model size of 6.2 MB, and an inference time of 8.9 ms/crop. The complete system achieved 96.8% end-to-end exact-match accuracy, a total model size of 18.4 MB, and a total inference time of 27.1 ms/image, demonstrating a practical accuracy–efficiency balance for the evaluated complex land-port scenes.