TY - GEN
T1 - Polarization Image Recognition Based on Cascade Deep Learning
AU - Li, Jinshan
AU - Chen, Hantang
AU - Ma, Xu
AU - Chen, Weili
N1 - Publisher Copyright:
© 2023 SPIE. All rights reserved.
PY - 2023
Y1 - 2023
N2 - Polarization imaging technology integrates the spatial and polarization information of the target scene, which can provide high-dimensional light field information to improve the ability of object detection and recognition. The polarization states of natural scenes can be characterized by the Stokes vector (S0 , S1 , S2 ), degree of polarization (DoP ) and angle of polarization (AoP ). In order to better understand and utilize the polarization characteristics, the observers need to recognize the feature maps of different polarization parameters. These images are sometimes hard to distinguish with naked eyes, especially for S1 and S2 images due to their similarity. This paper proposes a polarization image recognition method based on the cascade deep learning approach, which can improve the discrimination between S1 and S2 images, and achieve preferable recognition accuracy for different kinds of polarization images. We use two ResNet-50 networks successively to classify the polarization images. Firstly, a ResNet-50 network is used to recognize S0 , S12 , DoP and AoP images, where S12 means the union set of S1 and S2 images. Next, the Sobel operation is applied to enhance the discrimination of polarization characteristics between S1 and S2 images. After that, the second ResNet-50 network is used to separate the images of S1 and S2 . It shows that the proposed method outperforms some other comparative methods in terms of recognition accuracy.
AB - Polarization imaging technology integrates the spatial and polarization information of the target scene, which can provide high-dimensional light field information to improve the ability of object detection and recognition. The polarization states of natural scenes can be characterized by the Stokes vector (S0 , S1 , S2 ), degree of polarization (DoP ) and angle of polarization (AoP ). In order to better understand and utilize the polarization characteristics, the observers need to recognize the feature maps of different polarization parameters. These images are sometimes hard to distinguish with naked eyes, especially for S1 and S2 images due to their similarity. This paper proposes a polarization image recognition method based on the cascade deep learning approach, which can improve the discrimination between S1 and S2 images, and achieve preferable recognition accuracy for different kinds of polarization images. We use two ResNet-50 networks successively to classify the polarization images. Firstly, a ResNet-50 network is used to recognize S0 , S12 , DoP and AoP images, where S12 means the union set of S1 and S2 images. Next, the Sobel operation is applied to enhance the discrimination of polarization characteristics between S1 and S2 images. After that, the second ResNet-50 network is used to separate the images of S1 and S2 . It shows that the proposed method outperforms some other comparative methods in terms of recognition accuracy.
KW - Polarization imaging
KW - Sobel operation
KW - deep learning
KW - image recognition
UR - http://www.scopus.com/inward/record.url?scp=85180124488&partnerID=8YFLogxK
U2 - 10.1117/12.2689211
DO - 10.1117/12.2689211
M3 - Conference contribution
AN - SCOPUS:85180124488
T3 - Proceedings of SPIE - The International Society for Optical Engineering
BT - Third International Conference on Optics and Image Processing, ICOIP 2023
A2 - Li, Bingxiang
A2 - Ren, Chao
PB - SPIE
T2 - 3rd International Conference on Optics and Image Processing, ICOIP 2023
Y2 - 14 April 2023 through 16 April 2023
ER -