Lightweight model for detecting fresh corn cobs quality using improved SS-YOLOv8

This study aims to accurately and rapidly detect the fresh corn cobs in modern agriculture. A lightweight model was also proposed for the superior and inferior fresh corn cobs using the improved SS-YOLOv8. Among them, the YOLOv8 model was one of the five size versions in the YOLO series. Furthermore, the YOLOv8n also shared the most lightweight architecture and the high detection speed for real-time tasks. In addition, the high accuracy of the YOLOv8 was comparable to the advanced target detection. Compared with the traditional ones, the YOLOv8 model was the best choice to serve as the base model in the end-to-end detection with significantly higher speeds. Therefore, the quality of the corn cob was evaluated after detection. Firstly, the backbone of the network was improved for the feature extraction, where a large number of the kernels were independently and tightly arranged on the fresh corn cob. The loss of feature information was avoided to reduce the number using feature reuse lightweight network, ShuffleNetV2, the lightweight convolutional layer (SPD-Conv) with the fine-grained information, and the maximally-pooled convolutional layer (Conv_Maxpool). Secondly, a simple and parameter-free attention module (SimAM) was integrated into the backbone network of the feature extraction. The network was enhanced to extract the features from the defective corn cobs. Finally, the Wise-IoU (WIoU) was introduced as the regression loss function of bounding box for the YOLOv8n. The accuracy of the identification was further improved for the flaws of the corn cob. The convergence was also promoted during training. The experimental results show that the improved SS-YOLOv8 model effectively detected the qualified and unqualified fresh corn. The mean average precision (mAP) was achieved at 98.7%, which was 1.4 percentage points higher than the baseline model; The number of parameters and model size were only 1.99 M and 4.2 MB, respectively, which were reduced to 66.1% and 66.7% of the baseline model. Additionally, the various attention mechanisms (such as the SimAM, CBAM, and SE) were introduced into the feature extraction network. Among them, the Sim attention mechanism shared the best performance. The regression loss functions of the bounding box were assessed on the performance of the improved model. The WIoU outperformed GIoU, DIoU, and CIoU in the speed of convergence and the size of loss values. A systematic comparison was also made on the two-stage model, Faster R-CNN, the single-stage model, SSD, and the YOLOv5, v6, and v8 of the YOLO series. The results show that the precision, recall, and mAP of the SS-YOLOv8 model were much higher than those of the Faster R-CNN and SSD models, while slightly higher than those of the YOLOv8 and YOLOv5 models. In addition, the improved model was far better than the rest in terms of the computational parameters and size. In summary, the improved SS-YOLOv8 model shared significantly better performance in detection accuracy, the number of parameters, and model size, compared with the rest. As such, the SS-YOLOv8 model can be expected to realize the effective identification of the qualified and unqualified corn cobs with small weights and a number of computational parameters. The finding can provide a strong reference for the detection equipment of the maize grading.