A recognition algorithm applied to apple orchard picking robot

Abstract In complex orchard environments, the recognition accuracy of picking robots is often affected by lighting conditions, leaf occlusion, and overlapping apple images. To address the problem of inaccurate judgment of apple fruit maturity by picking robots in orchards, we conducted research on Red Fuji apples and proposed a comprehensive model for judging apple maturity. The model first accurately identifies apples through the instance segmentation algorithm Mask R-CNN, divides apple boundaries, and then combines apple image RGB and apple fruit shape proportion to comprehensively analyze apple Ripeness Ratio and fruit Aspect Ratio, and finally obtains apple maturity and judges whether to pick. We applied and tested the model through the MinneApple dataset, A total of 10,305 apples were detected, including 3,732 fully ripe apples, 4,074 half ripe apples, and 2,499 unripe apples, and the experimental results showed that the accuracy of the model was 91.4%, which can better meet the requirements of picking robot recognition. We hope that our model can provide ideas for research on different apple maturity judgments based on mechanical automation picking.