基于改进YOLO v3的自然场景下冬枣果实识别方法

为实现自然场景下冬枣果实的快速、精准识别，考虑到光线变化、枝叶遮挡、果实密集重叠等复杂因素，基于YOLO v3深度卷积神经网络提出了一种基于改进YOLO v3（YOLO v3-SE）的冬枣果实识别方法。YOLO v3-SE模型利用SE Net 的SE Block结构将特征层的特征权重校准为特征权值，强化了有效特征，弱化了低效或无效特征，提高了特征图的表现能力，从而提高了模型识别精度。YOLO v3-SE模型经过训练和比较，选取0.55作为置信度最优阈值用于冬枣果实检测，检测结果准确率P为88.71%、召回率R为83.80%、综合评价指标F为86.19%、平均检测精度为82.01%，与YOLO v3模型相比，F提升了2.38个百分点，mAP提升了4.78个百分点，检测速度无明显差异。为检验改进模型在冬枣园自然场景下的适应性，在光线不足、密集遮挡和冬枣不同成熟期的情况下对冬枣果实图像进行检测，并与YOLO v3模型的检测效果进行对比，结果表明，本文模型召回率提升了2.43~5.08个百分点，F提升了1.75~2.77个百分点，mAP提升了2.38~4.81个百分点，从而验证了本文模型的有效性。

Winter Jujube Fruit Recognition Method Based on Improved YOLO v3 under Natural Scene

Winter jujube fruit recognition is the key technology to realize automatic picking, fruit trees precision management and yield forecast in winter jujube orchard. The rapid and accurate recognition of winter jujube fruits in natural scene affects real-time operability of automatic picking and reliability of monitoring and prediction directly. According to the complex recognition conditions, such as dark light, backlighting, occlusion, and dense fruits in winter jujube orchards, YOLO v3-SE model embedded in SE Net was proposed based on YOLO v3. SE Net adaptively recalibrated channel-wise feature responses by explicitly modelling interdependencies between channels. It strengthened important and valid features, and weakened unimportant and invalid features to improve the performance of feature maps. The deep convolutional neural network built in the article was TensorFlow. After the YOLO v3-SE model was trained and its recognition effect was tested on test samples, and 0.55 was selected as the optimal confidence threshold for the final detection. The P, R, F and mAP were used to assess the differences between YOLO v3-SE and YOLO v3 models. Test results showed that the model proposed got significantly good results. The detection results had the P value of 88.71%, R value of 83.80%, F value of 86.19%, and mAP value of 82.01%. Compared with the results of YOLO v3, the F value and mAP value had an increase of 2.38 percentage points and 4.78 percentage points. Meanwhile, there was no significant difference in detection speed. The further experiments compared the test results of the proposed model and YOLO v3 in complex conditions. In the data sets of backlight and dark-light fruit, the F value and mAP value of the proposed model reached 83.10% and 76.58%. In the data sets of occlusion and dense fruit, the F value and mAP value of the proposed model were 85.02 % and 74.78%. In the data sets of white-ripe, crisp-ripe and full-ripe stage fruit, the F value and mAP value of the proposed model were 86.37%, 89.91%, 91.49%, and 81.18%, 85.15%, 87.49%, respectively. Compared with the original YOLO v3, the F value was increased by 1.75~2.77 percentage points, and the mAP value was increased by 2.38~4.81 percentage points. The detection performance was significantly improved. The above content verified the effectivity of the YOLO v3-SE model. The model proposed can provide a method for winter jujube automatic picking and orchard yield forecast.