Abstract:

When machine vision technology is used to automatically pick local characteristic forest fruit Lingwu long jujube, the naturally changing environment will greatly affect the accuracy of detection. In order to adapt to the changing natural environment, an area convolutional neural network（Faster R-CNN） ensemble learning model based on double loss function is proposed. Firstly, an image data set is established, including a training set and a test set. Secondly, a Faster R-CNN model is built based on the features. The RPN layer uses softmax as the basic classifier to obtain the region of interest. This feature map is combined with a layer of loss function. Use Large Marge Softmax Loss（L-softmax） and Angular Softmax Loss（A-softmax） to do loss calculations, and take various maximum values. Finally, load the training set for image training to get the detector, and use the trained detector. The result image is obtained from the test set, and the image detection is completed. Compared with the single loss function network, the standard ResNet101 and ResNet50 network structures, the image detection network method in this paper：the accuracy rate is 0.9826, the recall rate is 0.9213, and the average accuracy is 0.9.

Key words: Lingwu long jujube, Faster R-CNN, L-softmax, A-softmax, average precision

摘要：

基于机器视觉技术自动采摘地方特色林果灵武长枣时，自然变化的环境会极大地影响检测的准确率。为适应时时变化的自然环境，提出基于双损失函数的区域卷积神经网络（Faster R-CNN）集成学习模型。建立图像数据集，包括训练集和测试集；根据特征搭建Faster R-CNN模型，在RPN层利用softmax作为基础分类器得到感兴趣区域；以此为特征图再结合一层损失函数，分别利用Large Marge Softmax Loss（L-softmax）、Angular Softmax Loss（A-softmax）做损失计算，取各类最大值；加载训练集进行图像训练得到检测器，将已训练好的检测器通过测试集得出结果图像，完成图像检测。同单一损失函数网络、标准的ResNet101以及ResNet50网络结构进行对比，该图像检测网络方法的精确率为0.982 6，召回率为0.921 3，平均精度为0.9。

关键词: 灵武长枣, Faster R-CNN, L-softmax, A-softmax, 平均精度