Polarization colorization method based on deep learning

Abstract

The invention discloses a polarization colorization method based on deep learning. The polarization colorization method comprises the following steps: collecting images in four polarization directionsand color images corresponding to the same scene by adopting a polarization imaging system; calculating a polarization angle image and a polarization degree image of the acquired image, and calculating a polarization factor feature image corresponding to the scene; constructing a training set according to the color image of the scene and the corresponding polarization factor feature image; constructing a dual-flow architecture network of low-light polarization colorization based on deep learning; designing a loss function; and inputting the images in the training set into a deep neural network for training to obtain a polarization colorization model. According to the method, a colorized network with a double-flow architecture is adopted, global information and local information are fused,and sufficient extraction of image features is realized; human interference is avoided, the colorization effect of polarization imaging can be achieved, the observation habit of human eyes is met, and detail information of the scene is further enriched.

Description

technical field [0001] The invention relates to a polarization colorization method. Background technique [0002] Polarization information is information independent of amplitude, phase, and frequency. During the process of reflecting and radiating electromagnetic waves, an object will change its polarization state according to the surface properties of the object. The polarization information generated by different targets or different states of the same target is different. According to this characteristic, it can be effectively Distinguishing the target and background information, so as to realize the efficient exploration and recognition of the target, has far-reaching application prospects in both military and civilian fields. Therefore, the integration of polarization information with the existing imaging system can effectively overcome the problem of effective target information in the case of low contrast. Difficult to extract questions. At present, most of the exis...

AI Technical Summary

Problems solved by technology

The data-driven image coloring method is mainly to find a suitable reference image for the target coloring image, but the working process and results cannot be reused after each matching. When coloring a new image, it needs to be re-matched

Images