改进丰富卷积特征算法的液滴边缘检测模型  

Improved droplet edge detection model based on RCF algorithm

作　　者：王慧 曹召良 王军[2] WANG Hui;CAO Zhao-liang;WANG Jun(School of Physical Science and Technology,Suzhou University of Science and Technology,Suzhou 215009,China;School of Electronic and Information Engineering,Suzhou University of Science and Technology,Suzhou 215009,China)

机构地区：[1]苏州科技大学物理科学与技术学院,江苏苏州215009 [2]苏州科技大学电子与信息工程学院,江苏苏州215009

出　　处：《中国光学（中英文）》2024年第4期886-895,共10页Chinese Optics

基　　金：“十四五”江苏省重点学科资助(No.2021135);中国航天科技集团公司第八研究院产学研合作基金资助(No.SAST2020-025)。

摘　　要：液滴图像边缘的高精度提取是测量水接触角较为关键的一环,针对常规边缘提取方法噪声鲁棒性差、边缘提取不完整、精度低的问题,本文提出了一种改进丰富卷积特征(RCF)的液滴边缘检测模型。首先,在深度特征提取阶段引入特征融合模块,使用多个特征让模型更加鲁棒,减少过拟合的风险;其次,设计多感受野模块代替RCF后边的contact层,通过多个感受野来提取更多的语义信息,使边缘细节更加丰富;然后,在模型每一层之前引入高效通道注意力机制,增强模型对图像中重要特征的关注程度;最后,设计并引入MaxBlurPool下采样技术,减少计算量和参数量,提高平移不变性。在自制液滴数据集上的实验结果表明,本文模型的固定轮廓阈值(ODS)提高到0.816、单图像最佳阈值(OIS)提高到0.829、检测准确率高达90.17%,相较原模型提高了1.85个百分点,能够准确检测液滴边缘特征。Accurate droplet edge extraction is crucial for measuring water contact angle.To address issues like poor noise robustness,incomplete edge extraction,and low precision in conventional methods,we pro-pose an improved model for droplet edge detection based on Richer Convolutional Feature(RCF)algorithm.Firstly,a feature fusion module is introduced in the deep feature extraction stage to enhance model robust-ness and reduce overfitting risks.Secondly,a multi-receptive field module replaces the contact layer after RCF to extract more semantic information and enrich edge details.Thirdly,an efficient channel attention mechanism is introduced before each layer of the models to enhance focus on important features of the im-age.Lastly,the MaxBlurPool downsampling technique is designed and incorporated to reduce computation and parameter requirements while improving translation invariance.Experimental results on a self-made droplet dataset demonstrate that the proposed model achieves an ODS value of 0.816,an OIS value of 0.829,and a detection accuracy of up to 90.17%,which is an improvement of 1.85 percentage points compared to the original model.It can improve accuracy in droplet edge features detections.

关 键 词：深度学习 边缘检测 水接触角 特征融合 曲线拟合 

分 类 号：TP394.4[自动化与计算机技术—计算机应用技术] TH691.9[自动化与计算机技术—计算机科学与技术]