Purpose To propose an improved YOLOv4 approach for the recognition of mature tomato fruit, and to solve the prevalently inaccurate identification of tomato by picking intelligent robot, which was due to the occlusion by branches and leaves.
Methods The deep learning algorithm was employed to introduce lightweight modules such as Mobilenetv2, Mobilenetv3, and Ghostnet into the convolutional neural network YOLOv4 algorithm for the feature extraction and identification analysis of mature tomato fruits, with an occlusion degree of less than 35%, 35%-65%, and more than 65%.
Results The three lightweight modules of YOLOv4 could achieve a recognition rate of over 90% for tomato fruits with a shielding area of less than 65%. Among the three modules, the Ghostnet-YOLOv4 algorithm had the highest recognition rate and the most stable performance, with an average recognition rate of 94.41% and a detection speed of 0.012 s. The recognition rate of the algorithm for fruits covered by more than 65% was obviously better than that of the other two algorithms. The recognition rate of the three algorithms decreased as the occlusion degree increased.
Conclusion The Ghostnet-YOLOv4 algorithm can be applied to the identification of tomato fruits, with high recognition rate and fast speed, and it has obvious advantages in the identification of covered tomatoes. This study can provide better concepts and a technical foundation for picking strategy of robots.
DownLoad: