基于改进YOLOv3的自然场景中未成熟青苹果识别研究

Research on recognition of immature green apples in natural scenes based on improved YOLOv3

为研究适合当前生产实际的苹果自动套袋技术,实现真实环境中未成熟青苹果的精准识别,提出一种基于改进YOLOv3的自然场景中未成熟青苹果图像识别方法。首先,为提高含干扰因素图像中未成熟青苹果识别准确率,基于YOLOv3算法利用残差网络和多尺度特征融合检测小目标的思想,对YOLOv3特征提取网络进行改进与试验验证,利用尺寸为(104,104,128)的特征图代替原尺寸为(13,13,1024)的特征图作为输出,提出改进YOLOv3的未成熟青苹果目标检测模型,通过增大特征提取网络输出特征图尺寸,减小感受野尺寸,提高算法网络对图像中未成熟青苹果的捕捉能力与识别准确率。其次,设计不同算法、不同品种和不同环境下的识别对比试验并对结果进行对比分析。改进YOLOv3在整体数据集上的检测均值平均精确率mAP值和召回率R值分别为92.46%、87.6%,较原YOLOv3分别提高3.22%、14.57%,改进模型性能提升主要体现在检测正确目标数量的能力上;在含光照影响、重叠和遮挡影响图像测试集上改进YOLOv3的mAP值较原YOLOv3分别提高3.58%、2.74%。改进YOLOv3模型对整体数据集和含干扰因素图像测试集的检测准确率较高,检测正确目标的数量较多,抗干扰能力较强。

In order to study the automatic bagging technology for apples suitable for the current production practice and realize the accurate recognition of immature green apples in the real environment,this paper proposed an image recognition method of immature green apples in natural scenes based on the improved YOLOv3.Firstly,in order to improve the recognition accuracy of immature green apples in images containing interference factors,this paper was based on the idea that YOLOv3 algorithm utilized residual network and multi-scale feature fusion for detecting small targets,and improved and experimentally verified the YOLOv3 feature extraction network by utilizing the feature maps with dimensions of(104,104,128)instead of the original feature map with dimensions of(13,13,1024)as the output.The improved YOLOv3 target detection model for immature green apples was proposed,which improved the ability of the algorithm network to capture immature green apples in the image and the recognition accuracy by increasing the size of the output feature maps of the feature extraction network and decreasing the size of the receptive field.Secondly,this paper designed the recognition comparison test under different algorithms,different varieties and different environments,and compared and analyzed the results.The mean Average Precision and Recall of the improved YOLOv3 on the overall dataset were 92.46%and 87.6%,respectively,which were 3.22%and 14.57%higher than that of the original YOLOv3.The performance enhancement of the improved model was mainly reflected in the ability to detect the correct number of targets.The mean Average Precision of the improved YOLOv3 on the test set of images containing the effects of illumination,overlapping and occlusion effects was improved by 3.58%and 2.74%compared to the original YOLOv3,respectively.The improved YOLOv3 model had higher detection accuracy on the overall dataset and on the test set of images containing interference factors,higher number of correct targets detected,and better anti-interference ability.