Multi-class detection of cherry tomatoes using improved YOLOv4-Tiny

The rapid and accurate detection of cherry tomatoes is great significance to realizing automatic picking by robots. However, so far, are detected as only one class for picking. Fruits occluded branches or leaves pickable objects, which may cause damage the plant robot end-effector during This study proposed Feature Enhancement Network Block (FENB) based on YOLOv4-Tiny solve above problem. Firstly, according distribution characteristics strategies tomatoes, were divided into four classes in nighttime, daytime included not occluded, branches, fruits, leaves. Secondly, CSPNet structure with hybrid attention mechanism was used design FENB, pays more effective features different while retaining original features. Finally, (FEN) constructed FENB enhance feature extraction ability improve accuracy YOLOv4-Tiny. experimental results show that under confidence 0.5, average precision (AP) non-occluded, branch-occluded, fruit-occluded, leaf-occluded fruit over day test images 95.86%, 92.59%, 89.66%, 84.99%, respectively, 98.43%, 95.62%, 95.50%, 89.33% night images, respectively. mean Average Precision (mAP) set higher (94.72%) than (90.78%), both better YOLOv4 It cost 32.22 ms process a 416×416 image GPU. model size 39.34 MB. Therefore, can provide practical feasible method multi-class tomatoes. Keywords: deep learning, data augmentation, YOLOv4, occlusion, DOI: 10.25165/j.ijabe.20231602.7744 Citation: Zhang F, Chen Z J, Ali S, Yang N, Fu S L, Y K. Multi-class using improved Int J Agric & Biol Eng, 2023; 16(2): 225-231.