Chenglong Wang

Food defect detection is crucial for the automation of food production and processing. Potato surface defect detection remains challenging due to the irregular shape of potato individuals and various types of defects. This paper employs deep convolutional neural network (DCNN) models for potato surface defect detection. In particular, we applied transfer learning by fine-tuning a base model through three DCNN models—SSD Inception V2, RFCN ResNet101, and Faster RCNN ResNet101—on a self-developed dataset, and achieved an accuracy of 92.5%, 95.6%, and 98.7%, respectively. RFCN ResNet101 presented the best overall performance in detection speed and accuracy. It was selected as the final model for out-of-sample testing, further demonstrating the model’s ability to generalize.

The performance of fruit surface defect detection is easily affected by factors such as noisy background and foliage occlusion. In this study, we choose lychee as a fruit type to investigate its surface quality. Lychees are hard to preserve and have to be stored at low temperatures to keep fresh. Additionally, the surface of lychees is subject to scratches and cracks during harvesting/processing. To explore the feasibility of the automation of defective surface detection for lychees, we build a dataset with 3743 samples divided into three categories, namely, mature, defects, and rot. The original dataset suffers an imbalanced distribution issue. To address it, we adopt a transformer-based generative adversarial network (GAN) as a means of data augmentation that can effectively enhance the original training set with more and diverse samples to rebalance the three categories. In addition, we investigate three deep convolutional neural network (DCNN) models, including SSD-MobileNet V2, Faster RCNN-ResNet50, and Faster RCNN-Inception-ResNet V2, trained under different settings for an extensive comparison study. The results show that all three models demonstrate consistent performance gains in mean average precision (mAP), with the application of GAN-based augmentation. The rebalanced dataset also reduces the inter-category discrepancy, allowing a DCNN model to be trained equally across categories. In addition, the qualitative results show that models trained under the augmented setting can better identify the critical regions and the object boundary, leading to gains in mAP. Lastly, we conclude that the most cost-effective model, SSD-MobileNet V2, presents a comparable mAP (91.81%) and a superior inference speed (102 FPS), suitable for real-time detection in industrial-level applications.