Aerial image matching method based on HSI hash learning
Introduction
With the rapid development of aerial photography technology, aerial photography have been widely used in emergency rescue, environmental monitoring, digital city construction and military [1]. To obtain the physical and geometric information, aerial image matching plays a crucial role in the field of aerial photography. It is a technology that seeks for corresponding relations like contents, features, structure, texture and grayscale between images with the completion of similarity and consistency analysis. In recent years, UAV aerial photography technology has been developing rapidly and its superiority on clarity, dimension scale and area has increasingly been prominent. Nonetheless, with the continuous increase of aerial image data as well as the feature dimension, requirements for its matching rate and matching accuracy are becoming increasingly higher. Therefore, how to realize the feature point extraction and matching from high-resolution aerial images is of great research significance.
To realize the aerial image matching, the corresponding feature point descriptor needs to possess strong discrimination and robustness performance. Thus, floating-point descriptors like SIFT [2] and SURF [3] constructed by high-dimensional floating-point data have been put forward. These descriptors share rather positive robustness in terms of scale invariance and rotation invariance. Specific to variability of aerial images, the image information and matching precision can be well reserved. However, large memory space is required for traditional floating-point descriptors, and because of the high computational complexity, the real-time extraction and matching of high-resolution aerial images are severely restricted.
To solve problems like high computational complexity and high memory space occupation of floating-point descriptors, in-depth studies have been conducted on binary feature descriptors. According to obtaining methods, binary feature descriptors can mainly be divided into two categories. The first method is to transform the original floating-point data into binary string indirectly, such as the binary descriptor of random luminance difference quantification (IDQ) [4] , a horizontal or vertical mirror reflection invariant binary descriptor(MBR-SIFT) [5] and the multiple binary descriptor of entropy choice (EC) [6].etc. The second category is to select sampling points randomly within the neighborhood of the identified image feature point and then obtain directly by comparing the gray value, such as ORB [7], BRIEF [8] and BRISK [9]. Memory space occupation can be curtailed and the matching rate can be improved by using the binary descriptor, so it is more applicable to the storage and processing of high-dimensional data of aerial images. However, the discrimination and robustness performance of the feature descriptor itself is instable, so it fails to meet the precision requirements of aerial image matching. Therefore, how to further enhance the extraction and matching efficiency still requires further study.
In recent years, hash learning [10] technology has extensively used in data mining [11], image retrieval [12], [13], [14] and target identification [15]. Given that all the data is saved in the form of binary in the computer memory, so the computational burden and storage space will be significantly reduced by constructing a suitable hash function to generating the binary descriptors and calculating in the Hamming space. Nevertheless, hash learning technology still needs to be further studied. For instance, the traditional optimization strategy [16] may result in the encoding mapping error caused by optimizing all parameters simultaneously. In addition, in the quantification stage [17] of hash learning, spatial similarity information between feature points in the original space is often neglected.
Specific to the encoding mapping error and the quantization loss in the traditional hash learning, the HSI hash learning matching algorithm for aerial image matching is proposed. First, a HSI local multi-feature descriptor is proposed based on HSI space [18], [19], [20]. The HSI space is represented by a vertical intensity axis and the trajectory of color points in a plane perpendicular to the intensity axis. The model can eliminate the influence of the intensity component from the carried color information in the color image. It is an ideal tool for developing image processing algorithms based on color description. Then, the hash function is constructed according to HSI component images and the objective function is designed based on the hash information and similarity information. The projection matrix and bias threshold are determined respectively by the spectral relaxation and the problem of encoding mapping error caused by parameter computation can be relieved at the same time. Finally, the minimized loss distance function is defined in the hamming matching stage, so as to remain the spatial similarity between Euclidean space and Hamming space. While determining the threshold, the problem of neglecting similarity information can be reduced in the quantification stage.
Section snippets
Related work
Classical floating-point descriptors such as SIFT and SURF have been widely used in feature matching. With the surge of image data amount, the problems of floating-point descriptors like high computational complexity and high storage requirements have been shown, which are especially prominent in real-time tracking or image sequence matching. To reduce memory usage and increase computational efficiency, many improvement methods have been proposed. A new binary local descriptor Edge-SIFT [21]
HSI hash learning matching algorithm
As aerial images are rich in color information and hash learning is superior on storage and computational efficiency, an HSI hash learning matching algorithm is proposed in this paper. The HSI hash learning matching algorithm can be divided into three parts: construction of the local multi-feature descriptor, HSI hash learning and hamming matching based on minimized loss distance. Where, in the hash learning stage, the objective function is constructed for each component image with the hash
Experimental results and analysis
In this experiment, by choosing standard database and real UAV aerial images, the effectiveness of HSI hash learning matching algorithm has fully been verified. To prove the performance of feature descriptors of this paper, four standard databases, namely, Bike, Boat, Graggiti and Iguaze are selected from CV Datasets to check the robustness of the algorithm on blurriness variation, scale and rotation variation, visual angle variation and noise variation. In addition, by taking real aerial
Conclusions
To achieve fast matching of high-resolution aerial images, a multi-feature hash learning method with the integration of color information is brought forward. While guaranteeing the matching rate, the matching effect is improved. By combining the HSI color space, the ability to describe the aerial images has been greatly improved. Furthermore, through HSI hash learning, high-dimensional feature descriptive vector can be mapped into low-dimensional binary code, thereby realizing the fast matching
Acknowledgments
This work was supported by National Natural Science Foundation of China (61705019), Natural Science Foundation of the Jiangsu Higher Education Institutions of China (12KJA510001) and was funded by the Priority Academic Program Development of Jiangsu Higher Education Institution, China.
References (43)
- et al.
Kernelized sparse hashing for scalable image retrieval
Neurocomputing
(2016) - et al.
Local tri-directional patterns: a new texture feature descriptor for image retrieval
Digital Signal Process.
(2016) - et al.
Local derivative radial patterns: a new texture descriptor for content-based image retrieval
- et al.
BINK: Biological binary keypoint descriptor
Biosystems
(2017) - et al.(2016)
- et al.
Optimized SIFT image matching algorithm
- et al.
Research of image matching algorithm based on surf features
- et al.
Novel binary feature from intensity difference quantization
J. Image Graph.
(2014) - et al.
MBR-SIFT: a mirror reflected invariant feature descriptor using a binary representation for image matching
Plos One
(2017) - et al.
Multiple binary codes based on entropy selection
Jilin Daxue Xuebao
(2017)
An Improved ORB Algorithm of Extracting and Matching Features
BRIEF: binary robust independent elementary features
BRISK: binary robust invariant scalable keypoints
Learning binary hash codes for large-scale image search
Mach. Learn. Comput. Vis.
A hash-based algorithm for mining non-redundant asynchronous periodic patterns
Int. J. Bus. Intell. Data Min.
Unsupervised local feature hashing for image similarity search
IEEE Trans. Cybern.
Deep supervised hashing for fast image retrieval
Comput. Vis. Pattern Recognit.
Instance-aware hashing for multi-label image retrieval
IEEE Trans. Image Process.
A hash function algorithm based on variable parameter cascade chaos
J. Comput. Res. Dev.
Quantization-based hashing: a general framework for scalable image and video retrieval
Pattern Recognit.
Vehicle license plate detection and tilt correction based on HSI color model and SUSAN corner detector
Smart Comput. Rev.
Cited by (9)
A dichotomy color quantization algorithm for the HSI color space
2023, Scientific ReportsMultiscale Reflection Component Based Weakly Illuminated Nighttime Image Enhancement
2022, Circuits, Systems, and Signal ProcessingA Computer Vision System for Missing Tablets Detection
2021, Journal of Physics: Conference SeriesA Detection System for Rail Defects Based on Machine Vision
2021, Journal of Physics: Conference SeriesDetection of Cracks on Mobile Screen Based on Computer Vision
2020, Journal of Physics: Conference SeriesA fast, accurate and dense feature matching algorithm for aerial images
2020, Journal of Systems Engineering and Electronics