Fast and accurate recognition of the strawberries in greenhouse based on improved YOLOv4-Tiny model

Abstract: Detection of fruit images has been one of the most important steps for automatic picking robots. There are many factors that make strawberry detection difficult in an orchard, such as the complex background, fruit occlusion, and small target fruit. In this study, an improved detection model with the YOLOv4-Tiny was proposed to rapidly and accurately recognize the strawberries in the greenhouse for the high detection accuracy of small targets in an orchard. Firstly, the GhostNet lightweight network was adopted to replace the backbone CSPDarkNet53-tiny for the feature extraction, which significantly reduced the parameters and computation of the model. Convolution Block Attention Module (CBAM) with the spatial information was embedded into the Ghost Bottleneck module instead of the original Squeeze-and-Excitation (SE) attention, in order to improve the feature extraction capability. Secondly, the Spatial Pyramid Pooling (SPP) module was introduced in the neck network structure, and then to carry out the maximum pooling operation with three pooling kernels (5×5, 9×9, 13×13). Feature Pyramid Network (FPN) structure was adopted to improve the detection accuracy of small target strawberries. Finally, the Efficient Intersection over Union Loss (EIoU Loss) was used to separate the influence factors of aspect ratio, and then calculate the length and width of the target frame and anchor box. As such, the convergence speed was faster with higher regression accuracy than before. The original data set was collected, consisting of 841 strawberry images with a complex background in the greenhouse. Data enhancement was carried out for the training, the verification, and the test set, in order to improve the generalization ability of the model. The experimental results showed that the average accuracy of the improved YOLOv4-Tiny model in the test set was 92.62%, which was 5.77 percentage points higher than the original one. The average detection time of each image was 5.63ms, and the final model size was only 4.68MB. The average accuracies of the improved model were 9.11, 4.80, 2.26, 1.22, and 1.91 percentage points higher, while the F1 scores were 0.08, 0.05, 0.03, 0.01, and 0.03 higher than those of the SSD, CenterNet, YOLOv3, YOLOv4, and YOLOv5s target detection networks. The improved YOLOv4-Tiny model was much smaller than the SSD, CenterNet, YOLOv3, YOLOv4, where the network structure size was only 1/3 of YOLOv5s. The average detection speed of each image was only 5.63ms, which was reduced by 5.57, 7.14, 9.20, 15.99, and 2.16ms, respectively. Therefore, the improved YOLOv4-Tiny model can fully meet the requirements of high precision and real-time detection of strawberry fruits under the background of the orchard in the greenhouse. The finding can provide an effective way for the accurate detection of strawberry fruits in a complex environment using picking robots.