An Integrated Target Recognition Method Based on Improved Faster-RCNN for Apple Detection, Counting, Localization, and Quality Estimation

Aiming at the problems of the dense target distribution, poor positioning ability to pick robots, and inaccurate ripeness judgment in the picking orchard scene, this paper proposes an apple image recognition model with a high recognition rate, high speed, and high accuracy, which can effectively analyze the data of quantity, location, ripeness and quality estimation in the apple image. Firstly, the Faster R-CNN network is improved by introducing Efficient Channel Attention (ECA) and multi-scale fusion feature pyramid (FPN) for fruit detection and recognition localization. Then the distance transform-based watershed algorithm is used for image segmentation to fit the apple edge image while combining with the fitted circle determination algorithm to establish a mathematical model for apple volume estimation to calculate the quantity as well as the quality of apples. Finally, the apples are classified into four categories according to their ripeness, and the improved Faster R-CNN network is used to improve the ripeness detection effect, and the results show that the average fruit recognition accuracy of the improved method proposed in this paper is 95.42%, which significantly improves the accuracy of fruit detection.