Polarized hyperspectral low-altitude reconnaissance image typical target detection method

Abstract

The invention, in combination with the development of CNN frameworks, polarized hyperspectral imaging technology and dictionary pair learning, provides a novel polarized hyperspectral low-altitude reconnaissance image typical target detection method in a polarized hyperspectral low-altitude target detection simulation environment. The method is characterized by providing a dictionary pair driven CNN classifier for target detection by improving the mature Faster R-CNN in the CNN framework; using dictionary pair back propagation (DPBP) for the end-to-end learning of the dictionary pair classifier and the feature representation of CNN; using a sample weighting method to improve positioning performance; using multi-task loss for the joint training of a DPCL and boundary frame regression; introducing a polarized hyperspectral image into target detection and selecting three typical targets for testing so as to preliminarily verify the validity of the model and the sample. The method can be combined with different CNN frameworks, has high flexibility, can enhance the target-background contrast of the polarized hyperspectral image, reduces the complexity of the background to a certain extent, highlights the target, contributes to detection results, and has significant meaning to the improvement in the detection and recognition of polarized imaging targets.

Description

technical field [0001] The invention belongs to the fields of polarization imaging detection and computer vision, relates to a new target detection method, and is suitable for typical target detection of polarization hyperspectral low-altitude reconnaissance images. Background technique [0002] In recent years, the scale of UAVs used in battlefield reconnaissance and strikes has been increasing. The rapid and automatic identification of battlefield targets is one of the important performance indicators and development trends in the field of UAVs. While the quality of target images continues to improve , researching and improving the target detection algorithm can further improve the target detection efficiency of low-altitude platforms such as drones. [0003] Due to the development of deep convolutional neural network (CNN) and the continuous increase in the size of training data sets, object detection has produced breakthrough progress in recent years. State-of-the-art o...

AI Technical Summary

Problems solved by technology

Existing DDL methods suffer from two major deficiencies: first, methods that use conventional hand-held features (e.g., SIFT and HOG); second, involve heavy “L 0 " or " L 1 "Norm regularization to generate sparsely encoded vectors limits its use in scenarios with high feature dimensions and large amounts of data

Images