LES-YOLO: A lightweight pinecone detection algorithm based on improved YOLOv4-Tiny network

Pinecones are an important forest product that are often harvested manually. Automated pinecone detection and harvesting might increase the yield and resolve the problem of labor shortage. The quick and accurate detection of pinecones is one of the most critical technologies for ground picking equipment and very significant for reducing the leakage rate and increasing the yield. Currently, most solutions use deep learning to achieve pinecone recognition. However, deep learning has poor detection effect for small targets, large parameters, and slow operation speed; thus, there is scope for improving the pinecone recognition task. This study proposes a lightweight pinecone detection algorithm based on an improved You Only Look Once version 4 tiny (YOLOv4Tiny) algorithm. A lightweight Enhance ShuffleNet is used as a backbone to extract the pinecone features. A squeeze-and-excitation feature pyramid network is used to fuse multi-scale information, retaining only a 26 x 26 detection head for pinecone prediction. In addition, focal loss and alpha-complete intersection over union loss are used to reduce undetected and false detection rates. Data augmentation is used to improve the generalization of the algorithm, and channel compression reduces the number of model parameters. The average precision of the improved network reaches 95.33%, which is an increase of 3.56% compared with the original YOLOv4-Tiny network. Additionally, the parameters, floating point operations, and detection time of one picture are compressed to 12.22%, 17.35%, and 67.41% of the original network, respectively. Comparative experiments were conducted using the proposed method and common target- detection algorithms. The findings show that the average precision of the proposed method is higher than the corresponding values of the YOLOv4, YOLOv5x, YOLOv5s, and YOLOX-Tiny algorithms (95.17%, 95.27%, 92.10%, and 90.47%, respectively), with parameters decreasing to 1.13%, 0.83%, 10.20%, and 14.34%, respectively. The proposed algorithm detects 135 pictures per second to achieve real-time effects. Further, it can distinguish pinecones quickly and accurately in a near-color background and provide a more precise and lightweight solution for pickup and acquisition equipment. In addition, the design process of the proposed algorithm could be a reference for neural network design for agricultural applications.