CONTAINER DAMAGE DETECTION BASED ON DEEP RESIDUAL NETWORKS AND REAL-TIME OBJECT DETECTION ALGORITHMS

Abstract

Container damage detection is a critical component for enhancing automation and transport safety in modern smart ports. To address practical challenges such as complex background interference in port environments, a wide range of damage scales, and uneven data distribution, this study constructs a damage detection framework based on the deep residual network ResNet50 and the lightweight detection algorithm YOLOv8. The study first employs a Super-Resolution Generative Adversarial Network (SRGAN) to enhance image resolution, thereby improving the distinguishability of fine cracks and dents. By introducing a cross-entropy loss function and a cosine-annealed learning rate strategy, the model achieves robust learning for three major damage types: dents, perforations, and corrosion. Experimental results demonstrate that the ResNet50 model excels at feature extraction in complex backgrounds, achieving a classification accuracy of 76.99% on the validation set, with particularly outstanding performance in identifying dent-type damage with distinct features. Meanwhile, the YOLOv8 model, which incorporates an attention mechanism, exhibits significant advantages in inference speed and multi-object localization, with an average accuracy of approximately 0.85, effectively meeting real-time monitoring requirements. This study confirms the efficiency and practical value of deep learning technology in handling container image recognition tasks characterized by high dynamics and diverse operating conditions.