| 基于改进Faster R-CNN的马铃薯发芽与表面损伤检测方法 |
| |
| 引用本文： | 刘毅君,何亚凯,吴晓媚,王文杰,张丽娜,吕黄珍.基于改进Faster R-CNN的马铃薯发芽与表面损伤检测方法[J].农业机械学报,2024,55(1):371-378. |
| |
| 作者姓名： | 刘毅君 何亚凯 吴晓媚 王文杰 张丽娜 吕黄珍 |
| |
| 作者单位： | 中国农业机械化科学研究院集团有限公司 |
| |
| 基金项目： | 国家马铃薯产业技术体系项目（CARS-10-P28）、国有资本金项目（GZ202007）和农业农村部农产品产地初加工重点实验室开放项目（KLAPPP2022-01） |
| |
| 摘 要： | 发芽与表面损伤检测是鲜食马铃薯商品化的重要环节。针对鲜食马铃薯高通量分级分选过程中，高像素图像目标识别准确率低的问题，提出一种基于改进Faster R-CNN的商品马铃薯发芽与表面损伤检测方法。以Faster R-CNN为基础网络，将Faster R-CNN中的特征提取网络替换为残差网络ResNet50，设计了一种融合ResNet50的特征图金字塔网络（FPN），增加神经网络深度。采用模型对比试验、消融试验对本文模型与改进策略的有效性进行了试验验证分析，结果表明:改进模型的马铃薯检测平均精确率为98.89%，马铃薯发芽检测平均精确率为97.52%，马铃薯表面损伤检测平均精确率为92.94%，与Faster R-CNN模型相比，改进模型在检测识别时间和内存占用量不增加的前提下，马铃薯检测精确率下降0.04个百分点，马铃薯发芽检测平均精确率提升7.79个百分点，马铃薯表面损伤检测平均精确率提升34.54个百分点。改进后的模型可以实现对在高分辨率工业相机采集高像素图像条件下，商品马铃薯发芽与表面损伤的准确识别，为商品马铃薯快速分级分等工业化生产提供了方法支撑。
|
| 关 键 词： | 马铃薯 发芽 表面损伤 Faster R-CNN 高分辨率 |
| 收稿时间： | 2023/5/22 0:00:00 |
Potato Sprouting and Surface Damage Detection Method Based on Improved Faster R-CNN |
| |
| Institution: | Chinese Academy of Agricultural Mechanization Sciences Group Co.,Ltd. |
| |
| Abstract: | Germination and surface damage detection are crucial steps in the commercialization of fresh table potatoes. To address the low accuracy rate of high-pixel image object recognition in the high-throughput grading and sorting process of fresh table potatoes, a method for detecting potato sprouting and surface damage based on improved Faster R-CNN was proposed. Using Faster R-CNN as the baseline network, the feature extraction network in Faster R-CNN was replaced with the residual network (ResNet50), and a feature pyramid network (FPN) integrated with ResNet50 was designed to increase the depth of the neural network. A comparative model assessment and ablation studies were performed to empirically validate the efficacy of the proposed model and its modifications. The findings delineated that the enhanced algorithm demonstrated an average precision rate of 98.89% in identifying potatoes, 97.52% in discerning sprouting events, and 92.94% in recognizing surface defects. When benchmarked against the Faster R-CNN model, the adapted model incurred no additional computational time or memory overhead while manifesting a marginal decline of 0.04 percentage points in potato identification accuracy. Notably, it significantly elevated the average precision in detecting sprouting and surface imperfections by 7.79 percentage points and 34.54 percentage points, respectively. This augmented model was robust in high-resolution imaging environments facilitated by industrial-grade cameras and served as a cornerstone for the methodological advancement of automated grading and sorting processes in the commercial potato industry. |
| |
| Keywords: | potato sprouting surface damage Faster R-CNN high-resolution |
|
| 点击此处可从《农业机械学报》浏览原始摘要信息 |
| 点击此处可从《农业机械学报》下载免费的PDF全文 |
