A multi-scaled hierarchical structure model for multispectral image detection

• The 3D space and spectrum feature descriptor is constructed to expand the target detection to multispectral area.
• A composite template set is applied to detect targets with different sizes.
• A hierarchical structure model is established to improve the accuracy and efficiency.

In view of the complex information and low recognition efficiency of multispectral images, a multi-scaled hierarchical structure model based on locally adaptive regression kernels is proposed. First of all, we construct the 3D feature descriptors by fully using the spatial structure and the abundant spectral information of the template set and test images, in order to expand the target detection to the multispectral area. Then a template set with multi-scale and multi-pose is put forward to ensure that targets with different sizes can be identified by the proposed statistical principle of similarity. Meanwhile, the hierarchical structure model is established to reduce the time of recognition and to increase the accuracy. Experiment results show that this algorithm can gain good recognition effect in both colorful images and multispectral images of near infrared band, and it is also conductive to the improvements of recognition accuracy and efficiency.