Polarization imaging technology is an optical reconnaissance method based on the polarization characteristics of the target, which can simultaneously obtain a variety of information about the target. This information is very useful for visual tasks such as target detection. The traditional polarization parameter map as the input source for target detection has the problems of low accuracy and weak generalization ability. To solve the above problems, this paper builds a polarization parameter feature extraction network on the basis of Faster R-CNN, and adaptively generates polarization parameters. Perform feature extraction. It has been verified that the detection accuracy has been improved compared to the traditional network with polarization parameters as the input source.
Aiming at the problems of poor real-time performance of image target detection algorithm in airborne video and unused inter frame redundancy features, this paper proposes OFYOLOX method. For continuous airborne video, the key frame and non key frame are distinguished, and the YOLOX target detection algorithm without anchor box is used on the key frame, which improves the adaptability and simplicity of various targets; ON non-key frames optical flow estimation and feature transformation are used to obtain detection features, which achieves higher real-time requirements and the accuracy is 85.1%. It provides a reference for airborne video target detection.
The details and shape information of the target are effectively highlighted in the polarized image, which is more conducive to target detection. At present, the influence of different polarization parameters on the target detection task has not been studied in depth. There are problems that the ways of characterization of polarization parameter is so rich and varied that the polarization parameter is difficult to select, when we utilize the convolutional neural network (CNN) model to detect images obtained by polarimetric systems. In response to this problem, this paper uses the convolutional neural network (CNN) model to train a variety of polarized parametric images in many different input configuration for experimental comparison. Firstly, the sample data is acquired using a focal plane polarized camera, and the database is expanded using a data enhancement strategy to establish a polarized image data set. Then, different image input configurations are used as the training set, and the convolutional neural network (CNN) pre-training model is iteratively trained and fine-tuned to obtain the target detection model of the polarized image. Finally, in order to evaluate the performance of the model, the experimental trials are quantified by mean average precision (mAP) and processing time, and the influence of different polarization image input configurations on the detection model is analyzed. The experimental results show that compared with the model trained by single channel image configuration, the model trained by threechannel image configuration has better performance, but there is no obvious difference between models trained by different three-channel configurations.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.