Deep Learning-based Visual Recognition System for Fruit-Picking Robots

Fruit recognition plays an important role in automated picking, in order to improve the accuracy and real-time picking, this study uses deep learning methods to design a fruit-picking robot visual recognition system. First, multiple preprocessing methods are used to expand the sample data, and the images are proportionally cropped and scaled to make the image dataset more complete; subsequently, to improve the accuracy of image recognition, a multilayer Convolutional Neural Networks (CNN) is established; and the Adam algorithm is utilized to train the model parameters several times to determine the optimal hyperparameters, thus overcoming the defects of the multilayer neural network that has a local optimal search. The experimental results show that compared with the random forest method, this system has the characteristics of high-speed recognition and high accuracy in fruit picking, and can quickly and accurately recognize fruit images, with the recognition speed of a single image taking only 0.081 seconds, and the recognition accuracy reaching more than 97.35%. The method has important theoretical and application value and provides an effective means for automatic fruit recognition.