半导体光电, 2019, 40 (1): 108, 网络出版: 2019-03-25  

基于深度学习的单步目标检测器特征增强算法

Feature Enhancement Algorithm for Single Shot Detector Based on Deep Learning
作者单位
1 中国科学院光电技术研究所, 成都 610000
2 中国船舶工业系统工程研究院, 北京 100036
摘要
针对基于深度学习的目标检测网络模型多采用级联的卷积网络结构进行特征提取, 没有很好地利用多尺度特征融合的信息, 以及卷积往往采用方形卷积核而没有提取出具备方向性的特征等问题, 提出了一种特征提取模块, 采用不同大小形状的卷积核结合异性卷积核并行提取特征, 并进行融合。该类结构相比于级联网络更能提取并融合目标的多尺度特征, 同时提取具有方向性的特征。提出的特征增强型单步目标检测器(Feature Enhanced Single Shot Detector, FESSD)网络基于单步目标检测器(Single Shot Detector, SSD), 修改了网络结构、加入特征提取模块并采用多层特征融合, 在VOC0712数据集上大大提高了检测准确率。
Abstract
The model based on deep learning for object detection usually uses cascaded convolutional structure, and does not make good use of fusion for multi-scale features. Meanwhile, it often uses a square convolution kernel which cannot extract directional features. In order to solve these two problems, a new feature extraction module was proposed, for which, the feature extraction block uses the convolution kernels with different sizes and shapes combining with the dilated convolution kernel to extract features and fuse them. Such a structure is more capable of extracting both multi-scale features and directional features than a cascade network. It is proposed to modify the network structure of FESSD based on single shot detector (SSD) network, add the feature extraction module and adopt multi-layer feature fusion, which greatly improves the detection accuracy on the VOC0712 data set.

孙元辉, 徐智勇, 张建林, 许涛. 基于深度学习的单步目标检测器特征增强算法[J]. 半导体光电, 2019, 40(1): 108. SUN Yuanhui, XU Zhiyong, ZHANG Jianlin, XU Tao. Feature Enhancement Algorithm for Single Shot Detector Based on Deep Learning[J]. Semiconductor Optoelectronics, 2019, 40(1): 108.

关于本站 Cookie 的使用提示

中国光学期刊网使用基于 cookie 的技术来更好地为您提供各项服务,点击此处了解我们的隐私策略。 如您需继续使用本网站,请您授权我们使用本地 cookie 来保存部分信息。
全站搜索
您最值得信赖的光电行业旗舰网络服务平台!