Apple detection algorithm based on G-YOLO

In view of the complex operating environment of apple picking robot, a G-YOLO apple detection algorithm based on lightweight YOLOv4 was proposed, which solved the problems of slow detection speed and large memory consumption of YOLOv4 model. The G-YOLO detection network model draws lessons fromthe idea of GhostNet. The traditional convolution operation is carried out in two steps, and the lightweight operation is used to enhance and extract features, so astoreduce the calculation amount of model parameters, and the detection ability is improved by introducing attention mechanism into the backbone network. The PANet network is optimized by the BI-FPN of EfficientDet.Based on the SqueezeNet principle, the SPPNet structure was optimized by FireModule and SPPF, and the model parameters were further reduced by deep separable convolution. Finally, the size of the model was only33.36MB, and the memory occupancy was reduced by 86 degrees A.The results showed that the mAP of G-YOLO was97.05%, which was 0.5% higher than YOLOv4.G-yolo improves the detection speed on GPU by about 24%, on CPU by 333%, and on Raspberry PI by 406.96%. Itis suitable for small embedded devices with weak computing power such as Apple picking robots.