A Lightweight Deep Learning Framework for Chinese and Western Pastry Texture Evaluation and Defect Detection

Introduction: Texture evaluation is a critical aspect of food quality control, particularly in the industrial production of Chinese and Western pastries. Traditional manual inspection methods suffer from low efficiency, high labor costs, and subjectivity, making automated defect detection a pressing need. Methods: This study proposes a lightweight deep learning framework that integrates three novel modules: Heterogeneous Depthwise-Diverse Fusion (HDDF), Channel-Adaptive Interaction (CAI), and Convolutional Block Attention Module (CBAM). These modules are designed to improve feature extraction, inter-channel interaction, and spatial alignment, respectively. The model was trained and tested on two comprehensive pastry defect datasets covering both Chinese filled dumplings and Western puff pastries, which incorporate diverse defect types and challenging imaging conditions. Results: Experimental results demonstrate that the proposed method outperforms existing lightweight detection models. On the Chinese pastry dataset, our model achieves a mean Average Precision at IoU 0.5 ([email protected]) of 92.7% with an inference speed of 238 FPS on an RTX 4090 GPU, surpassing MobileNetV3-YOLO, ShuffleNetV2-CBAM, and FDD-Net. On the Western pastry dataset, the model attains an [email protected] of 90.1%, outperforming BakeInspect, Defect- GAN-SSD, and Coordinate-Attention-YOLO. Discussion: The proposed framework demonstrates robust performance in accurately detecting fine-grained defects such as surface cracks, filling exposure, and shape deformation, even under motion blur and illumination shifts. Its modular design and high-speed performance make it suitable for real-time deployment in industrial food production lines. The study also highlights the limitations of conventional lightweight models in handling spatial-channel misalignment and redundant computation. Conclusion: This research presents a high-performance, lightweight deep learning framework for real-time detection of pastry defects. The integration of HDDF, CAI, and CBAM modules contributes to both computational efficiency and detection accuracy. The framework holds strong potential for practical deployment in automated food quality control systems and provides a foundation for future enhancements involving self-supervised learning and multimodal defect analysis.