Pitaya detection using an improved lightweight Faster R-CNN based on MobileNetV3 in densely planted pitaya orchards

Abstract

Accurate and rapid fruit detection was very important for robot picking precisely, so the large model size and slow detection speed of the detection algorithm are problems that need to be solved urgently. An improved lightweight Faster R-CNN based on MobileNetV3 was proposed in this paper, which was used to detect fruits on Ori_RGB and Rb_RGB image datasets that collected by RGB-D camera in densely planted commercial pitaya orchards. On the Rb_RGB image datasets, the detection AP of 0.929 and 0.898 were obtained using MobileNetv3_large_FRCNN and MobileNetv3_small_FRCNN, which decreased 1.38% and 4.67% than that using VGG16_FRCNN respectively, and the detection time was 35.4 and 18.8 ms per image, which decreased 46.5% and 71.6% than that using VGG16_FRCNN respectively. On the Ori_RGB image datasets, the detection AP of 0.911 and 0.856 were obtained using MobileNetv3_large_FRCNN and MobileNetv3_small_FRCNN, which decreased 2.15% and 8.06% than that using VGG16_FRCNN respectively, and the detection time was 35.2 and 19.5 ms per image, which decreased 47.2% and 70.8% than that using VGG16_FRCNN respectively. Weight sizes of MobileNetv3_large_FRCNN and MobileNetv3_small_FRCNN were 3.19%, 1.15% of that of VGG16_FRCNN respectively. The detection AP values on the Rb_RGB image test set using three networks than that on Ori_RGB image test set increased 1.98%, 4.91%, and 1.18%, but image type had no significant effect on AP. The improved lightweight Faster R-CNN based on MobilenetV3 is expected to deploy to the embedded system of the fruit picking robot to detect pitaya, which would promote the development of robot picking technology.