Especially, the proposed technique was able to simultaneously detect individual kiwifruit in clusters, which provides a promise for accurate yield mapping and multi-arm robotic harvesting.

Comparing against the conventional methods, it suggested that the proposed method has higher recognition rate and faster speed. The technique took 0.274 s to process each image (for images with 2352×1568 pixels) and only 5.0 ms on average to detect a fruit. Overall, the model reached a recognition rate of 92.3%. The test results showed that the recognition ratio of occluded fruit, overlapping fruit, adjacent fruit and separated fruit were 82.5%, 85.6%, 94.3% and 96.7%, respectively. Finally, another 100 images of kiwifruit canopies in the field environment (including 5,918 fruits) were used for testing the network. The average precision of the Faster R-CNN-based kiwifruit detector was 89.3%. A faster R-CNN was trained end-to-end by using back-propagation and stochastic gradient descent techniques with Zeiler and Fergus network (ZFNet). Sub-images were used as network training and validation samples. The PASCAL VOC Challenge performance has been significantly boosted by the prevalently CNN-based pipelines like Faster R-CNN. Under different lighting conditions, 2,100 sub-images with 784×784 pixels were prepared by random sub-sampling from 700 field captured images with a pixel resolution of 2352×1568 pixels. A kiwifruit detection system for field images was developed based on the deep convolutional neural network, which has a good robustness against the subjectivity and limitation of the features selected artificially.

S and r images