摘要
鉴于传统自适应阈值分割方法在检测中无法根据镜框粗细、颜色深浅以及拍摄背景复杂程度自适应双阈值门限。现采用深度学习中的CenterNet模型,通过MobileNetV3中的bneck结构堆叠,组成CenterNet模型的主干特征提取部分,从用户图片中获得镜框和眼睛的类别信息、位置信息,最后通过相对位置关系计算出测量要求的相应尺寸。该过程无需对图像进行阈值分割,实现了非结构化场景中镜框尺寸的快速测量。实验结果表明,该方法不受镜框粗细以及拍摄背景的影响,检测精度高达98.01%,每秒处理图片的数量(frames per second, FPS)达到15帧/秒,训练后的权重文件从152 MB减小到47 MB,并有较强的泛化能力。
Since the traditional adaptive threshold segmentation method can not adapt the double threshold according to the thickness of the frame,the depth of the color and the complexity of the shooting background.Now we use the CenterNet model in deep learning and form the main feature extraction part of CenterNet model by stacking the bneck structure in MobileNetV3.The category information and position information of glasses and eyes are obtained from the pictures of users wearing glasses,and the corresponding size of frame measurement requirements is calculated through the relationship between position information.The whole process does not need the threshold segmentation of the image.Finally,we realize the fast measurement of the frame size in the unstructured scene.The experimental results show that the measurement results are not affected by the frame thickness and the shooting background,the detection accuracy is as high as 98.01%,FPS reaches 15 frames/s,the weight file after training is reduced from 152 MB to 47 MB,and has the stronger generalization ability.
作者
陈奕帆
付东翔
傅迎华
陈杰
CHEN Yifan;FU Dongxiang;FU Yinghua;CHEN Jie(School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China)
出处
《控制工程》
CSCD
北大核心
2023年第11期1971-1978,共8页
Control Engineering of China
基金
国家自然科学基金资助项目(61605114)。
