Potato Bud Detection with Improved Faster R-CNN

Highlights A series of strategies including feature concatenation, tweaks of ResNet50, and modification of the default anchors with the chaos optimization-based k -means algorithm were proposed to improve the detection performance of the original Faster R-CNN. The improved Faster R-CNN achieved an average precision of 97.71%, which is 5.98% higher than that of the original Faster R-CNN and 14.38% higher than that of YOLOv2. The improved Faster R-CNN greatly boosted the detection performance for potato buds without incurring any noticeable additional computational overhead. Abstract. This article proposes an improved Faster R-CNN model to achieve better detection performance for potato buds, with the goal of preparing for the automated cutting of seed potatoes. Detection results of Faster R-CNNs with eight pretrained networks were compared, and ResNet50 was adopted as the backbone network in Faster R-CNN. On this basis, three model strategies, including feature concatenation, tweaks of ResNet50, and modification of the default anchors with the chaos optimization-based k-means algorithm, were proposed to improve the detection performance for potato buds. Experimental results on the test set demonstrated that the improved Faster R-CNN achieved an average precision (AP) of 97.71%, which is 5.98% higher than that of the original Faster R-CNN and 14.38% higher than that of YOLOv2. In addition, the average running time per image with the improved Faster R-CNN was 0.166 s, the same as that of the original Faster R-CNN. In other words, the improved Faster R-CNN greatly boosted the detection performance for potato buds without incurring any noticeable additional computational overhead, thus satisfying the requirements for real-time processing. Consequently, the improved Faster R-CNN can provide a solid foundation for the automated cutting of seed potatoes. Keywords: Chaos optimization-based k-means, Default anchors, Faster R-CNN, Feature concatenation, Potato bud detection, Tweaks of ResNet50.