Research on Fruit Detection Algorithms for Picking Robots in Complex Environments

The most suitable RGB space was selected by analyzing 11080 apple images in RGB, HSI and Lab color spaces. Image preprocessing was carried out using Gaussian filtering thereby improving the image quality. Each apple in the preprocessed image was labelled with its ripeness category using LabelImg software, and training and test sets were created. A ripeness estimation model based on YOLOv5s deep learning was established, the weights were optimised using the SGD algorithm, and the YOLOv5s detection algorithm was used to classify apple ripeness and predict the number of apples. After testing, the accuracy of the model reaches 0.91, the recall rate reaches 0.89, and the IOU reaches 0.74, which has good accuracy.