An efficient detection method for litchi fruits in a natural environment based on improved YOLOv7-Litchi

• Published in: Computers and electronics in agriculture 2024-02, Vol.217, p.108605, Article 108605
• Main Authors: Li, Can, Lin, Jiaquan, Li, Zhao, Mai, Chaodong, Jiang, Runpeng, Li, Jun
• Format: Article
• Language: English
• Subjects: Attention mechanism; Lightweight network; Litchi; Target detection; YOLOV7-Litchi network
• ISSN: 0168-1699; 1872-7107
• DOI: 10.1016/j.compag.2023.108605

•An improved YOLOV7-Litchi for litchi detection in complex environments was proposed.•A dataset of litchi fruits in complex orchard environments was established.•The proposed method achieves satisfactory results in lightweight, speed and accuracy. The identification and detection of small fruit targets combined with mechanized harvesting can effectively solve the problems of high labor intensity, high labor cost and low harvesting efficiency in modern litchi harvesting. The orchard environment background of litchi fruits is relatively complex, which makes it difficult for current object detection algorithms to identify small targets and densely distributed litchi fruits. To improve the detection of litchi fruit in complex environments, this paper proposed an improved YOLOv7-Litchi detection algorithm. Compared with the YOLOv7 network, the improved YOLOv7-Litchi integrates ELAN-L and ELAN-A modules based on a lightweight ELAN on a backbone network and replaces the ELAN modules, which make the network structure lightweight. The CNeB module in ConvNeXt was added to the backbone network, and the inverted bottleneck structure and larger convolution kernel size improved the performance of the network by reducing information loss and amplifying the global interactive representation. In addition, a lightweight CBAM module integrating channel attention and spatial attention was added and embedded into the ELAN-A module to better utilize and extract features. Finally, the addition of the transformer module CoT with multihead self-attention enhanced the visual representation of the neural network by tightly combining the contextual information of the input matrix. The experimental results showed that the YOLOv7-Litchi network proposed in this paper provides good detection ability for litchi fruits characterized by a dense distribution and mutual occlusion in complex backgrounds, with a recall rate of 95.90 %, an accuracy of 94.60 %, an mAP of 98.60 %, and an average detection time of 8.6 ms. These results represent improvements of 0.9 %, 1.6 %, 0.6 % and 2.4 ms compared with the recall, accuracy, mAP and average detection time of the YOLOv7 model, respectively. The proposed algorithm has the advantages of high precision, fast detection speed and strong robustness, thus providing a theoretical basis for the mechanized picking of litchis.