Fruit detection plays a vital role in robotic harvesting platforms. However, natural scene attributes such as illumination variation, branch and leaf occlusion, clusters of tomatoes, shading, etc. and double scene including image augmentation and natural scene have made fruit detection a difficult task. An improved YOLOv3 model termed as Tomato detection models, which includes YOLODenseNet and YOLOMixNet was applied to solve these problems. YOLODenseNet incorporated DenseNet backbone, while the backbone of YOLOMixNet combined DarkNet and DenseNet. With the incorporation of spatial pyramid pooling (SPP), feature pyramid network (FPN), complete (CIoU) loss and Mish activation function into both models, the tested accuracy of YOLODenseNet at 98.3 % and YOLOMixNet at 98.4 % on natural scene performed better than YOLOv3 at 96.1 % and YOLOv4 at 97.6 %, but not with YOLOv4 under the double scene. Furthermore, the obtained detection speed of YOLOMixNet at 47.4FPS was noted to be in close par with the YOLOv4 at 48.9FPS. Finally, the Tomato detection models showed reliability, better generalization, and a high prospect for real − time harvesting robots.

水果检测在机器人收割平台中起着至关重要的作用。但是，自然场景属性（例如光照变化，分支和叶子遮挡，西红柿簇，阴影等）以及包括图像增强和自然场景在内的双重场景使水果检测成为一项艰巨的任务。解决这些问题的方法是使用一种改进的YOLOv3模型（称为“番茄检测模型”），其中包括YOLODenseNet和YOLOMixNet。YOLODenseNet合并了DenseNet主干，而YOLOMixNet的主干合并了DarkNet和DenseNet。通过将空间金字塔池（SPP），特征金字塔网络（FPN），完整（CIoU）损失和Mish激活功能结合到两个模型中，YOLODenseNet和YOLOMixNet在自然场景下的测试精度分别为98.3％和98.4％，优于YOLOv3为96.1％，YOLOv4为97.6％，但不能在双重场景下使用YOLOv4。此外，注意到在47.4FPS时获得的YOLOMixNet检测速度与在48.9FPS时的YOLOv4接近。最后，番茄检测模型显示出可靠性，更好的通用性以及实时收获机器人的高前景。