Real-time detection of navel orange fruits in the natural environment based on deep learning

Abstact: Deep learning is widely used in intelligent picking, but the adverse effects of different environmental scenes on target detection and recognition are crucial to picking robots’ accurate and efficient work. First, the data set needed for the experiment was manually created. The data set selected 925 navel orange images, including 290 backlit sunny days, 310 forward light, and 325 cloudy days. The training and test sets were divided into 8:2. Then, we studied the detection of navel orange based on the improved model of single-stage target detection network PP-YOLO. Used the backbone network ResNet with deformable convolution to extract image features and combined with FPN (feature pyramid network) for feature fusion to achieve multi-scale detection. The K-means clustering algorithm clustered the appropriate Anchor size for the target navel orange, which reduced the training time and the confidence error of the prediction frame. Loaded the pre-trained model and compared the model performance with the original PP-YOLO, YOLO-v4, YOLO-v3, and Faster RCNN network. Analyzed the Loss curve and AP curve of the training log, the task of detecting navel oranges under sunny, sunny, and cloudy conditions was realized. Finally, the improved PP-YOLO detection accuracy was 90.81%, 92.46%, and 94.31%, and the recognition efficiency reached 72.3 fps, 73.71 fps, and 74.9 fps, respectively. The model performance is better than the other four, with better robustness. CCS CONCEPTS • Computing methodologies∼Artificial intelligence∼Computer vision∼Computer vision tasks∼Vision for robotics