With the rapid development of the food industry, packaging damage has become a significant threat to product quality and food safety. Traditional manual inspection methods suffer from low efficiency and high misjudgment rates, failing to meet the dual demands for real-time performance and accuracy in modern production lines. This paper proposes an adaptive damage detection algorithm for food packaging based on lightweight convolutional neural networks (CNNS). It integrates experimental platforms to achieve an automated detection process from image acquisition, model training to real-time inference. The algorithm introduces a multi-scale convolution feature extraction mechanism in its design, enhancing the ability to identify various types of damage such as tiny cracks, creases, and tears. Additionally, it improves the robustness and generalization performance of the model through data augmentation and regularization techniques. Experimental results show that the model achieves rapid inference within 0.13 s per image while maintaining high detection accuracy, making it suitable for complex backgrounds and multi-material packaging scenarios. This study provides an efficient and deployable technical approach for intelligent quality inspection of food packaging, with good potential for educational application and practical implementation.