Lightweight classification of lotus seedpod quality using improved YOLOv3-Tiny model

An accurate and efficient classification of quality is one of the most important steps to realize the automatic processing of lotus after harvest. However, it is still lacking in the quality classification of lotus after harvest. In this study, the classification of lotus quality was established, according to the evaluation scheme for white lotus quality published by the Chinese Association of Traditional Chinese Medicine. An improved YOLOv3-Tiny (You Only Look Once version 3-Tiny) model was also proposed for quality classification. Firstly, a vertical camera was deployed on an image acquisition platform for lotus seedpod, according to the principles of image recognition. The images of lotus seedpods were collected under three lighting conditions. A comprehensive test dataset was then constructed to augment using three image augmentation techniques: rotation, color switching, and affine transformation. The K-means clustering was also employed to optimize the scales of the prior anchor boxes, in order to enhance the regression accuracy. Ultimately, six scales were generated for the prior anchor boxes, namely (14, 18), (19, 21), (20, 27), (36, 44), (42, 49), and (50, 56). The accuracy rate was achieved at 85.48%. Subsequently, the SPP (Spatial Pyramid Pooling) module was added to extract feature information, according to the original YOLOv3-Tiny backbone network. Finally, the parameter evolution module of YOLOv3-Tiny was used to evolve a set of appropriate hyperparameters for the model. Ablation experiments were carried out to verify the effectiveness of data augmentation and redesign the prior anchor box scales. The backbone network of feature extraction was then optimized to improve the performance in the related hyperparameter evolution. The ablation experiments showed that the recognition precision increased by 4.95% after data augmentation. The precision of recognition increased by 0.76% after the K-Means algorithm to reunite classes. Spatial Pyramid Pooling (SPP) module was added to improve the precision of recognition by 1.12%. The hyperparameter evolution module was added to improve the precision of recognition by 93.1%, indicating the high overall robustness of the model. A test was then conducted to verify the effectiveness of the improved YOLOv3-Tiny model on the quality classification of lotus seedpod. A comparison was also made between the original YOLOv3-Tiny network model and Faster R-CNN. The results show that the improved YOLOv3-Tiny model achieved a Mean Average Precision (mAP) of 96.80%, a Precision (P) of 93.10%, and a Recall (R) of 94.60%. Compared with the original YOLOv3-Tiny model, the mAP, the Precision (P), and the Recall (R) increased by 12.49%, 14.77%, and 11.59%, respectively. The number of frames transmitted per second reached 25 Hz, which was 1.24 times that of the Faster R-CNN network model. Ablation experiments showed that the Precision (P), Recall (R), and mAP of the improved YOLOv3-Tiny model increased by 4.95%, 4.52%, and 5.61%, respectively. Once the prior anchor box scale was redesigned, the Precision (P), Recall (R), and mAP of the network model were 0.76%, 0.10%, and 0.25% higher than those of the model with the original prior anchor scale. The SPP module was introduced to increase the Precision (P), Recall (R), and mAP by 1.12%, 0.10%, and 0.23%, respectively. The Precision (P), Recall (R), and mAP increased by 7.94%, 6.87%, and 6.40%, respectively, after hyperparameter evolution. Ablation experiments verified that data augmentation, redesigning the prior anchor box scale, the YOLOv3-Tiny feature extraction backbone network, and related hyperparameter evolution effectively enhanced the detection performance of the network model for all levels of lotus after harvest. The data show that the improved algorithm shared a better recognition for lotus seeds, and fully met the requirement for real-time detection. The finding can also provide a technical reference for the classification of lotus quality.