摘要
目前番茄采摘主要依靠人工,实现番茄产业机械化和智能化刻不容缓,而番茄检测是最基础也最重要的一步。针对该问题,提出一种基于改进Mask RCNN的番茄检测算法。该算法选择ResNet50和FPN作为主干网络,提出一种新型RoI提取器,并在算法模型中使用空洞卷积(Atrous)。通过Labelme自制番茄数据集,将改进算法在自制数据集上进行训练和测试。结果表明,与Faster RCNN和Mask RCNN模型相比,改进后的模型AP值分别提高了5.5%和4.7%,AR值分别提升了6.8%和4.6%。该算法不仅提高了番茄的识别准确率,还更好地实现了实例分割。
At present, tomato picking mainly relies on manual labor, so it is urgent to realize the mechanization and intelligence of the tomato industry, and tomato detection is the most basic and most important step. In response to this, propose a tomato detection algorithm based on improved Mask RCNN. The algorithm selects ResNet50 and FPN as the backbone network, proposes a novel RoI extractor, and uses atrous convolution(Atrous) in the algorithm model. Through the Labelme self-made tomato data set, the improved algorithm will be trained and tested on the self-made data set. Compared with the Faster RCNN and Mask RCNN models, the improved model also increases the AP value by 5.5% and 4.7%, respectively, and the AR value that’s an increase of 6.8% and 4.6%, respectively.The results show that it not only improves the recognition accuracy of tomatoes, but also better achieves instance segmentation.
作者
高倩
诸德宏
封浩
GAO Qian;ZHU De-hong;FENG Hao(School of Electrical and Information Engineering,Jiangsu University,Zhenjiang 212003,China)
出处
《软件导刊》
2023年第2期75-80,共6页
Software Guide
