To improve the accuracy of fruit detection,four types of fruit images,which is pineapple,snake fruit,dragon fruit,and banana,were collected to construct a fruit detection dataset.The AutoAssign model is applied to fruit detection and improve-ments is proposed based on this model.Multi-scale training strategy is adopted,and RegNet network is used to replace ResNet50 network which is the original backbone feature extraction network.The mean average precision of the improved model reached 86.8%,which is 5.4 percentage points higher than the original AutoAssign model.And the frame per second of the proposed model achieves 34.6 img·s-1.Compared with Faster R-CNN,RetinaNet,and FCOS models,the mean average precision of this model has increased by 6.0,5.7,and 5.0 percentage,respectively.The proposed model significantly improves the accuracy of fruit detection,and the detection speed is acceptable,which has certain reference significance for fruit detection.
fruit detectionAutoAssignmulti-scale trainingRegNet network
NETL
NSTL
万方数据
