基于深度学习的苹果树侧视图果实识别

Fruit identification with apple tree side view based on deep learning

关键词:  苹果树  侧面果实  广域场景  深度卷积神经网络  Faster-RCNN  多目标识别

Key words:  apple tree  side view  wide complex scene  deep convolution neural  Faster-RCNN  multi-target identification

摘要:

【目的】传统果树侧面果实识别方法精度难以满足实际果实识别的需求，研究深度学习方法对提高苹果树侧面果实识别精度、增强模型对苹果复杂生长环境的适应性和泛化性具有重要意义。【方法】文章提出基于深度卷积神经网络对广域复杂背景环境下的侧面苹果特征进行学习的方法，完成苹果树侧面果实多目标识别任务。【结果】在广域复杂场景下，基于VGG16 为特征提取网络的Faster-RCNN 多目标检测模型在果实多目标检测任务中，识别精度达到91%，单幅影像识别时间约为1.4 s，相较于ResNet50 作为特征提取层的目标检测模型识别精度提高4%；在相同影像数据下，模型识别精度的主要影响因素是遮挡，导致模型漏判果实数量较多，VGG16 在不同程度遮挡区域的漏判率比ResNet 低6%。【结论】基于VGG16 卷积神经网络果树侧视图果实识别算法对广域复杂场景下的果实提取效果较好，特别是在具有遮挡情况下的识别结果更优，能够为果园产量估算提供一定的借鉴。

Abstract:

［Purpose］The accuracy of the traditional fruit identification method from fruit tree side is difficult to meet the actual fruit identification and counting requirements. Therefore，in order to improve the fruit identification accuracy of the apple tree and enhance the adaptability and generalization of the model to the complex growth environment of apple planting，this paper proposes a deep learning method.［Method］The convolutional neural network learns the characteristics of the apples from tree side view in the wide-area complex background environment，and completes the multi-target identification and counting task of the apple trees.［Result］The correlation identification results show that in the wide complex scene，the Faster-RCNN multi-target detection model based on VGG16 as the feature extraction network has an identification accuracy of 91% in the multi-target detection task，and the single image identification time is about 1.4 s. Compared with ResNet50 as the feature extraction layer，the target detection model has a high identification accuracy of 4%. Under the same image data，the factors affecting the identification accuracy of the model are mainly occlusion，resulting in a large number of fruits missing from the model，and VGG16 is in different extents. The miss rate is 6% lower than ResNet.［Conclusion］The experimental results show that the proposed algorithm has better fruit extraction effect in wide-area complex scenes，especially in the case of occlusion， which can provide a reference for orchard yield estimation.