基于FasterR-CNN的蓝莓冠层果实检测识别分析

朱旭 马淏 姬江涛 金鑫 赵凯旋 张开

摘要：【目的】基于Faster R-CNN模型對不同成熟度蓝莓果实进行精准识别分类，为浆果类果实的自动化采摘、产量预估等提供技术支撑。【方法】选取成熟果、半成熟果、未成熟果剪切图像各4000幅和8000幅背景图像作为训练集，1000幅原始图像用于验证集开展试验，改进Faster R-CNN算法，设计一种对背景干扰、果实遮挡等因素具有良好鲁棒性和准确率的蓝莓果实识别模型，模型通过卷积神经网络（CNN）、区域候选网络（RPN）、感兴趣区域池化（ROI Pooling）和分类网络来实现蓝莓图像背景消除及果实识别并与DPM算法进行对比。【结果】以WOA算法优化的训练参数作为参考，在蓝莓数据集上训练网络模型。在分析P-R曲线后计算F发现，Faster R-CNN算法在成熟果、半成熟果和未成熟果上的F值分别为95.48%、95.59%和94.70%，与DPM算法相比平均高10.00%。在对3类蓝莓果实的识别精度方面，Faster R-CNN同样有着优秀的识别效果。对成熟果、半成熟果和未成熟果的识别准确率分别为97.00%、95.00%和92.00%，平均识别准确率为94.67%，比DPM算法高20.00%左右。该网络模型在高精度的识别效果下，对于蓝莓果实的平均识别速度依然达0.25 s/幅，能满足实时在线识别的需求。【建议】获取多角度、复杂环境下的图像用来提高模型识别率;利用迁移学习建立蓝莓识别模型;数据集样本扩充并分类。

关键词： 蓝莓;冠层果实;Faster R-CNN;精准识别;产量预估;不同成熟度

Abstract：【Objective】With the help of Faster R-CNN model， accurate detection， identification and classification of blueberry canopy fruits under natural environmental conditions could be realized，which provided reference for automatic picking and yield prediction of berries. 【Method】Each 4000 images and each 8000 background images of ripe fruit， half-ripe fruit， immature fruit were as the training set， 1000 original images were used to verify test set， optimized Faster R-CNN algorithm was used to design a kind of blueberry fruit recognition model for fruits at different maturity levels which had good robustness and accuracy under the condition of background interference and fruit block. The model realized image background elimination of blueberries and recognition through convolution neural network（CNN）， regional candidate networks（RPN）， interested in regional pooling（ROI pooling） and the classification of network. 【Result】The trai-ning parameters optimized by WOA algorithm were used as a reference， and the network model was trained on the blueberry data set. After analyzing P-R curve and calculating F value， it was found that the F values of Faster R-CNN algorithm on ripe fruit， half-ripe fruit and immature fruit were 95.48%， 95.59% and 94.70%， which was 10.00% higher than that of DPM on average. In terms of the recognition accuracy of three kinds of blueberry fruits， Faster R-CNN also had excellent recognition effect. The recognition accuracy of ripe fruit， half-ripe fruit and immature fruit was 97.00%， 95.00% and 92.00% respectively， and the average recognition accuracy was 94.67%， which was about 20.00% higher than DPM algorithm on average. Under the high-precision recognition effect， the average recognition speed of the network model for blueberry fruit still reached 0.25 s/amplitude， which could meet the demand of real-time online recognition. 【Conclusion】The network model can be improved in the following aspects：to obtain images in multi-angle and complex environment to improve the model recognition rate;establish a blueberry recognition model with migration learning;data set samples should be expanded and classified.

Key words：blueberry; canopy fruit; Faster R-CNN; accurate identification; production forecast; different maturity

0 引言

【研究意义】我国自2000年开始规模化种植蓝莓，2018年国内蓝莓栽培总面积达55344 ha，产量达18.4238万t，分别占全世界蓝莓栽培总面积和总产量的40.996%和28.129%（张珊珊等，2015;李亚东等，2016;刘绍金和卢炜乐，2018）。……