【摘要】 针对自然场景下百香果果实密集,生长环境相对复杂,大规模种植带来人工识别、采摘和估计产量困难等问题,提出了改进Faster R-CNN的百香果目标检测算法,实现无遮挡、遮挡、重叠和背景四类果实自动检测和产量预测。该方法首先采用ResNet网络融合FPN对百香果进行多尺度特征提取;然后采用RPN网络提取ROI区域;最后,通过全连接层实现百香果分类和检测。经测试集验证,该方法在4类情况下检测的平均精确率达到87.98%,其平均准确率和召回率分别达到90.79%和90.47%,每幅图片的检测时间在0.178 s左右;产量估算中,其准确率为96.80%。结果表明,基于FPN+ResNet-101特征提取的Faster R-CNN目标检测算法能应用于自然场景下百香果的快速、准确检测和产量估算。更多还原

【Abstract】 Passion fruit is dense in orchards and the growing environment is relatively complex, which makes it difficult to identify, pick and estimate the yield for passion fruits. In this paper, the colour image of the passion fruit in the natural growth environment is taken as the research object, and the fruit is divided into four categories: unobstructed, occluded, overlapping fruits and background, respectively, and the improved faster R-CNN target detection algorithm is proposed for the automatic detection and yield prediction of passion fruit under the deep learning framework.Firstly, residual network fusion FPN is used to extract the multi-scale features of passion fruit, then the ROI region is extracted by RPN network, and finally achieve the classification and detection of passion fruit through full connection layer. In the test set, its best mean average precision reached 87.98%, the average precision and average recall rate reached 90.79% and 90.47%, respectively. The detection time of each picture is about 0.178 s. This algorithm achieves the 96.80% of precision rate in yield estimation.Therefore, the faster R-CNN target detection algorithm based on FPN+ResNet-101 feature extraction can be applied to the fast and accurate detection of passion fruitand yield estimation.更多还原