Optimized Faster R-CNNfor Fruit Detection of StrawberryHarvesting Robot

摘要

Due to the advantage of obtaining image features voluntarily, the deep-learning methods have been used in target detection of images in recent years. The strawberry harvesting robots had low accuracy under the influences of illumination, fruits overlap, occlusion or other natural factors. To eliminate the limitations above, an optimized Faster RCNN (Faster Regional Convolution Neural Network) algorithm was proposed in this paper. The backbone network with 13 convolutional layers was adopted to extract the feature maps. To generate the rotatable regions of interest, a rotation angle parameter a was introduced into the Region Proposal Network (RPN). The rotatable regions of interest were trained to predict the rotation bounding box of the target, which provided the target posture information. The fruit detection results of 169 test images showed that the average detection precision rate was 94.54% and the recall rate was 94.01%. Compared with four traditional methods, the method proposed in this paper demonstrated improved universality and robustness in a non-structural environment, particularly for overlapping and hidden fruits, and those under varying illumination.