This paper proposes a Faster R-CNN based detector for kiwi fruit detection. In order to alleviate the shortcomings of insufficient datasets and to avoid retraining a network for a long time, the idea of transfer learn...This paper proposes a Faster R-CNN based detector for kiwi fruit detection. In order to alleviate the shortcomings of insufficient datasets and to avoid retraining a network for a long time, the idea of transfer learning (Transfer learning) is used to train the kiwi dataset. Firstly, the kiwi data set in the natural environment was collected and prepared, and the different algorithms were compared with Faster R-CNN in the kiwi dataset. The experimental results also showed that Faster R-CNN is the best kiwi dataset for this paper. Different classification networks are then used as the backbone feature extraction networks for the Faster R-CNN algorithm. Finally, combined with the idea of transfer learning, the pre-trained model weights are first loaded on the COCO dataset, and then the training skills of Fine-tune are adopted to freeze the parameters of different parts of the model to train the kiwi dataset. The final experimental results show that the Faster R-CNN algorithm with backbone network VGG19 is better suitable on the kiwi dataset, and the detection accuracy is 2.69% higher than the original algorithm.展开更多
文摘This paper proposes a Faster R-CNN based detector for kiwi fruit detection. In order to alleviate the shortcomings of insufficient datasets and to avoid retraining a network for a long time, the idea of transfer learning (Transfer learning) is used to train the kiwi dataset. Firstly, the kiwi data set in the natural environment was collected and prepared, and the different algorithms were compared with Faster R-CNN in the kiwi dataset. The experimental results also showed that Faster R-CNN is the best kiwi dataset for this paper. Different classification networks are then used as the backbone feature extraction networks for the Faster R-CNN algorithm. Finally, combined with the idea of transfer learning, the pre-trained model weights are first loaded on the COCO dataset, and then the training skills of Fine-tune are adopted to freeze the parameters of different parts of the model to train the kiwi dataset. The final experimental results show that the Faster R-CNN algorithm with backbone network VGG19 is better suitable on the kiwi dataset, and the detection accuracy is 2.69% higher than the original algorithm.
