Sun Rui, Sun Xiaobing, Chen Feinan, Song Qiang, Pan Hao
Appl Opt. 2020 Jan 1;59(1):151-155. doi: 10.1364/AO.59.000151.
Polarimetric imaging detection is a relatively new and largely undeveloped field. Although convolutional neural networks (CNNs) have achieved great success in two-dimensional (2D) normal intensity images in the field of target detection, traditional CNN methods have not been widely applied to optical polarimetric images, and they cannot take full advantage of the connection between different polarimetric images. To solve this problem, three-dimensional (3D) convolutions are adopted to consider the relationship between S0, S1, and S2 images as a third dimension. Based on the 3D convolutions, a CNN with 3D and 2D convolutional layers is introduced to further improve the success rate of target detection with limited polarimetric images. The evaluations in different natural backgrounds reveal that the proposed method achieves higher detection accuracy than that of two traditional methods for comparison.
偏振成像检测是一个相对较新且在很大程度上尚未开发的领域。尽管卷积神经网络（CNN）在目标检测领域的二维（2D）正常强度图像中取得了巨大成功，但传统的CNN方法尚未广泛应用于光学偏振图像，并且它们无法充分利用不同偏振图像之间的联系。为了解决这个问题，采用三维（3D）卷积将S0、S1和S2图像之间的关系视为第三维。基于3D卷积，引入了具有3D和2D卷积层的CNN，以在有限的偏振图像下进一步提高目标检测的成功率。在不同自然背景下的评估表明，所提出的方法比两种用于比较的传统方法具有更高的检测准确率。
