基于改进U-Net++的燕窝杂质检测方法
Impurity detection in bird's nest based on improved U-Net++
韦龙星 1宁萌 1蔡礼扬 1凤鹏锦 1陈义亮1
作者信息
- 1. 江南大学 机械工程学院,江苏无锡 214122;江南大学 江苏省食品先进制造装备技术重点实验室,江苏无锡 214122
- 折叠
摘要
为满足燕窝领域的杂质自动化检测,实现对燕窝中羽绒杂质的快速精准分割,提出一种应用于燕窝领域的两阶段杂质检测算法.第一阶段,基于U-Net++模型引入注意力模块,以抑制因燕窝强度不均所引起的图像分割不精准和密集卷积造成的干扰噪声;第二阶段,针对特征提取输出的概率张量,通过二值掩膜以实现对燕窝杂质和非杂质区域的精确分类.采集并预处理燕窝图像,通过消融试验对比分析杂质检测算法与U-Net,U-Net++,传统图像方法在同等条件下的测试集数据.试验表明,杂质检测算法的F1 系数为94.80%,较其他3种算法分别提高2.78%,1.12%,20.71%,召回率为97.90%,精确率为91.89%,整体检测结果优于对比算法.研究为燕窝杂质检测提供一种新的思路.
Abstract
In order to meet the requirements of automated impurity detection in the field of bird's nest to achieve rapid and accurate segmentation of down impurities in bird's nest.A two-stage impurity detection algorithm applied in the field of bird's nest was proposed.The first stage introduces an attention gate(AG)based on the U-Net++model to suppress interference noise caused by inaccurate image segmentation and dense convolution due to uneven bird's nest strength.In the second stage,the probability tensor of feature extraction output is used to achieve precise classification of bird's nest impurity and non-impurity regions through binary masks.bird's nest images are collected and preprocessed,and the test set data of impurity detection algorithms are analyzed and compared with U-Net,U-Net++and traditional image methods under the same conditions through ablation experiments.The experiment shows that the F1 coefficient of the impurity detection algorithm is 94.80%,which is 2.78%,1.12%,20.71%higher than the three algorithms,with a recall rate of 97.90%and an accuracy rate of 91.89%.The overall detection results are better than the comparison algorithms.The study provides a new approach for impurity detection in bird's nest.
关键词
图像处理/U-Net++/AG/杂质检测/机器视觉Key words
image processing/U-Net++/AG/impurity detection/machine vision引用本文复制引用
基金项目
国家重点研发计划(2022YFD2100304)
国家自然科学基金(51705201)
出版年
2024
国家科技期刊平台
NSTL
万方数据