基于YOLOv3网络的自然环境下青苹果检测与识别

Detection and recognition of green apple in natural environment based on YOLOv3 network

针对自然环境下青苹果目标与树叶颜色相似导致检测与识别困难的问题,提出一种基于YOLOv3网络的青苹果检测与识别方法。利用YOLOv3网络检测出图像中的青苹果目标区域,对目标区域进行HSV和YUV颜色空间分量下的阈值分割,选取青苹果目标提取效果较好的H、V和Y、U分量下的结果,通过形态学运算去除不连通的小区域得到青苹果目标。实验结果表明:在单个果实、两个果实和多个果实图像中,H、V和Y、U分量下青苹果目标提取的真阳性率均值为90.12%,假阳性率为5.74%,其中YUV颜色空间下Y分量的青苹果目标识别效果最好,真阳性率均值为93.93%。

Regarding the difficulty of detecting and recognizing the color similarity between the green apple target and the leaves in natural environment,a green apple detection and recognition method based on YOLOv3 network is proposed.The YOLOv3 network is used to detect the green apple target area in image.The threshold segmentation under HSV and YUV color space components in the target area is performed,and the H,V and Y,U components with better extraction effect of the green apple target are selected.Morphological operations remove the small areas that are not connected to get the green apple target.In the images of single fruit,two fruits and multi-fruits,the experimental results show that the true average positive rate of green apple target extraction under H,V,Y,and U components is90.12%,and the false average positive rate is 5.74%.Meanwhile,the Y component green apple target recognition in the YUV color space is the best,and the true average positive rate is 93.93%.