Multi-scale Hierarchical Target Detection Method Based on Spatial-Spectral Structure Constraints

Abstract

The present invention proposes a multi-scale layered target detection method based on space-spectral structure constraints. The method includes the following steps: constructing a multi-scale, multi-view and multi-attitude template set according to the target to be detected; calculating the binary neighbor feature matrix and the adaptive local kernel regression feature matrix of the template set, and performing de-redundancy optimization; The image to be tested is reduced to 1 / 4 of the original image, and the background noise in the scene is removed; the neighbor feature matrix is ​​calculated, and the similarity is compared with the neighbor feature matrix of the template set according to the cosine similarity criterion, so as to determine the range T of the target 1 ;Reduce the image to be tested to 1 / 2 of the original image, and calculate T 1 The adaptive local kernel regression feature matrix of the region, and compare the similarity with the template set to determine the target region T 2 ;Restore the image to be tested to the original image size, calculate T 2 The neighborhood feature matrix of the area is compared again to determine the final position of the target. The invention effectively shortens the detection time while improving the detection precision.

Description

technical field [0001] The invention belongs to the field of target recognition, and the specific method relates to a combination of multi-dimensional window features in a layered structure model. Background technique [0002] Object recognition is an important topic in the field of computer vision, which includes classification and detection. Object classification is to classify a given object into several known categories, while the latter is to extract the object from the image under test. Traditional object classification is a classifier model that relies on statistical learning and requires a large number of training samples. This training process is slow and prone to parameter overfitting. As a result, training-free object detection algorithms have been developed in recent years. [0003] In 2003, Document 1 (D. Comaniciu, V. Ramesh, P. Meer, Kernel-based object tracking, Pattern Analysis and Machine Intelligence, IEEE Transactions on, 25 (2003) 564-577.) proposed t...

AI Technical Summary

Problems solved by technology

[0004] Document 6 (Xue T, Han J, Zhang Y, et al.Aneighboring structure reconstructedmatching algorithm based on LARK features[J].Infrared Physics&Technology,2015,73:8-18.) Although considered from the perspective of non-negative linear reconstruction Neighborhood structure, but the structural relationship of most objects in the natural environment is nonlinear, and linear reconstruction is bound to fail to achieve good results; at the same time, the essence of NRSM is only simple feature fusion, neighborhood reconstruction and LARK features at the window scale There is no regional inclusion relationship formed on the above, but the image blocks are processed separately

Images