Abstract: In order to realize automatic fruit picking in natural environment, there are problems caused by environment, obstacles and other factors, resulting in low detection accuracy and generalization of target fruit and other practical problems. Taking apples, oranges and bananas as research objects, a Single Shot Detector（SSD） improved model ?is?proposed based on deep learning. The classical SSD adopts the multi-scale feature fusion method to extract different scale features from different layers of network for prediction, but does not use enough low-level features to make detection effect of small objects poor. VGG16?input?model?is?replaced?for?classical?SSD?training?with?ResNet?101, ?and?the?Feature?Pyramid?Network （FPN） structure?is?adopted?to?fuse?upper-level?features?with?low-level?features. Experiment results show that the average detection accuracy of the improved SSD300 and SSD512 fruit detection model is 83.05% and 84.24%. There is also a small increase after the data is enhanced, which is suitable for accurate detection of fruits in natural environments.

Key words: deep learning, target detection, Single Shot Detector（SSD） model, ResNet-101 model, ?Feature?Pyramid Network（FPN）

摘要： 为实现自然环境下水果自动化采摘存在受环境和障碍物等因素造成的问题，导致目标水果检测准确率不高，泛化性不强等实际问题，以苹果、橘子、香蕉三种水果作为研究对象，提出一种基于深度学习的SSD（Single Shot Detector）改进模型。经典SSD采用多尺度特征融合的方式，从网络不同层抽取不同尺度的特征做预测，但是没有用到足够低层的特征，使得小物体的检测效果较差。通过将经典SSD训练使用的VGG16输入模型替换为ResNet-101，利用特征金字塔网络（FPN）结构将高层特征通过上采样和低层特征做融合。实验表明，改进的SSD300和SSD512水果检测模型的平均检测精度为83.05%和84.24%，经数据增强后精度也有所提升，适合于自然环境下水果的精确检测。

关键词: 深度学习, 目标检测, SSD模型, ResNet-101模型, 特征金字塔网络（FPN）