Reparameterization of a Lightweight Network for Strawberry Flowers Detection

An improved YOLO v4 with lightweight parameters was proposed to detect strawberry flowers in real-time. The improvement was accomplished through two strategies. Firstly, the backbone network was lightweight designed by replacing standard convolution with grouped convolution, under the Visual Geometry Group (VGG) and re-parameterization VGG (Rep-VGG) frameworks. Secondly, the backbone network was fused into a single branch after the training phase through reparameterization method. The experiment results shown that the improved algorithm was better than the original version in detection precision and speed, as well as memory footprint. Furthermore, the results also outperform the excellent object detection algorithms, including Faster R-CNN, SSD, YOLOX-s and EfficientDet. The proposed algorithm was suitable to be deployed on embedded platforms with limited computing power for strawberry flowers detection.