摘要
为解决微孔板液位识别中材质透光差异和液面反光导致识别精度不足的问题,提出了一种改进U-Net模型(FDEGA-Unet)。该模型采用VGG16网络作为编码器;解码器使用频率自适应空洞卷积增强上下文信息捕获能力;跳跃连接处加入边缘引导注意力模块强化液位边界特征提取,以提升边界分割精度。在自建微孔板液位数据集上的实验结果表明,FDEGA-Unet的分割性能显著优于现有语义分割模型,关键指标,包括平均交并比、平均像素准确率和准确率,分别达到97.46%、98.87%和99.56%,且在不同光照环境下表现出更强的鲁棒性。该模型有效克服了材质干扰和光照变化的影响,实现了高精度、高效率的微孔板液位识别,为药物筛选、核酸检测等医学检测的自动化分析提供了技术支持。
To address the issues of insufficient recognition accuracy in microplate liquid level recognition caused by material translucency variations and liquid surface reflection,an improved U-Net model(FDEGA-Unet)was proposed.The model employs VGG16 network as its encoder.The decoder utilizes frequency-adaptive dilated convolution to enhance the capabilities of capturing contextual information,while an edge-guided attention module incorporated into the skip connections strengthens the extraction of liquid boundary features to improve segmentation precision.Experimental results on a self-constructed microplate liquid level dataset demonstrate that the segmentation performance of FDEGA-Unet significantly outperforms existing semantic segmentation models.Key metrics,including mean intersection over union(mIoU),mean pixel accuracy(mPA),and accuracy,reach 97.46%,98.87%,and 99.56%respectively.The model also exhibits greater robustness under varying illumination conditions.FDEGA-Unet effectively mitigates interference from material properties and illumination fluctuations,enabling high-precision,efficient liquid level recognition,providing technical support for automated analysis in medical detection applications including drug screening and nucleic acid testing.
作者
左浩瑜
张晓青
郭阳宽
ZUO Haoyu;ZHANG Xiaoqing;GUO Yangkuan(College of Instrument Science and Opto-electronics Engineering,Beijing Information Science&Technology University,Beijing 100192,China)
出处
《北京信息科技大学学报(自然科学版)》
2025年第5期77-84,共8页
Journal of Beijing Information Science and Technology University(Science and Technology Edition)
关键词
液位识别
语义分割
频率自适应空洞卷积
边缘引导
liquid level recognition
semantic segmentation
frequency-adaptive dilated convolution
edge-guidance
