Using Deep Convolutional Features for MultiTemporal Remote Sensing Image Registration
1S. S. ALEGAVI, 2R.R.SEDAMKAR
1PhD. Research Scholar, Electronics & Telecommunication Engineering,
2Professor & Dean Academic, Computer Engineering Thakur College of Engineering & Technology
Mumbai, India
Abstract
Image Registration is a fundamental part of image classification algorithm, for example, two or more images taken from various sensors or from different points of view at different times and environmental conditions may lead to degradation of accuracy factor during classification of any traditional algorithms. Image Registration will help us to minimize tedious task to select features in such a case. Specific examples of systems in which image registration is an important component include matching a target with a real-time image of a target scene, monitoring global land use[4], geographical maps using satellite images[5], matching stereo images to recover autonomous navigation shape, and aligning pictures from different diagnostic medical modalities etc. In this paper, we propose a flexible and dynamic algorithm, which is a fusion of DeepIRDI (Image Registration Using Deep Convolutional Features and Dynamic Inlier Selection) with MS – MA (multi-scale multi-angle) and CNN (Convolutional Neural Network)[1]
architecture for both rigid and non-rigid feature matching of remote sensing images. The performance of this technique is determined by using various datasets, and the results are evaluated using the hybrid DeepIRDI algorithm. For getting better and optimal result we will modify existing DeepIRDI with some hybrid features selection technique and performance analysis will be done using NVDIA GPU.
Keywords: Convolution neural networks (CNNs), Hyperspectral imager (HSI), Synthetic aperture radar (SAR), MS-MA BT (multi-scale multi-angle breaking ties), DeepIRDI (Image Registration Using Deep Convolutional Features and Dynamic Inlier Selection)
1. Introduction
Synthetic aperture radar (SAR) imaging systems now produce gigabytes of SAR pictures. Image registration is a technique of aligning two or more images captured at separate moments or from distinct points of view. For many remote sensing analysis functions, such as image fusion, change detection, image mosaic, etc., it is a basic preliminary step. The image registration generally includes the following four steps (1) extraction of the feature, (2) matching of the feature, (3) estimation of the transformation model and (4) resampling of the image [6][8]. Extraction of features plays a vital role in the registration of images as it chooses which sort of function to use for matching images.
Since deep learning can automatically learn the characteristics from images as a fully data-driven system, it has lately been introduced to remote image registration. Most methods of registration images are using deep learning techniques [5][6]. However, there are presently no public training datasets for recording remote sensing images, and training samples must be labelled manually before implementing the Deep Learning method. Due to the diverse nature of remote sensing data, i.e. images captured at different resolution, at
distinct moments (or by distinct modalities), the establishment of big public training datasets for image registration will be a big challenge and laborious task.
Classification and change detection of remote sensed images has been effectively implemented in many fields. Remote-sensed image processing, however, includes a few pre-processing algorithms in addition to classification and change detection; it is also highly dependent on the algorithm used [4]. The remote sensing community is therefore always dedicated to developing different methods of remote sensing to improve the efficiency of elements such as pre-processing, segmentation, and classification. Deep learning (DL) algorithms have been used for many years in the remote sensing community [10].
As illustrated in figure 1, our model involves the phase of noise removal, convolution, feature concatenation and classification. We use Gabor and ICA filter to denoise the hyperspectral images in the phase of noise removal [3].
This paper introduces a convolution neural network (CNN) based on multitemporal remote sensing image registration technique with two main contributions:
(i) we have used CNN to generate multi-scale and multi-angle feature descriptors[2] and (ii) we design a gradually growing range of inliers to improve feature point registration &
robustness. Extensive experiments on matching features and image registration are conducted over a satellite data set of multi-temporal image.
This paper is arranged as follows. Our proposed algorithm is described in Section 2. In Section 3, we experimentally evaluate the efficiency of our technique with different parameters. Finally, conclusions are made in Section 4.
2. Implementation
We suggest a hybrid DeepIRDI model and CNN algorithm with noise removal phase, feature extraction phase, matching phase, convolution phase, concatenation phase and identification phase. This method utilizes the SAR images with noise, multi-scale and multi-angle characteristics. The Derived results will be analyzed in quantitative and qualitative domain. Quantitative analysis gives results in time and size. In this system we have used CUDA GPU since it is fast as compare to normal processor and it requires less time for processing.
Our proposed system uses hybrid DeepIRDI algorithm [1] in fusion mode along with other algorithms. It gives result in multi-scale and multi-angle variants. In this system we have used SAR image dataset. On that image the Rescaling and angle theta will be applied after noise filtering. Dataset is provided in resolution 512* 512, resolution 256 * 256 and resolution 64* 64. Furthermore, these images are rotated by
angles respectively and a new dataset is created with all the possible permutations and combinations to test accuracy factor of our proposed system. All resolution datasets are fused together with angle. For knowing the multi-scale deep characteristics[2], we have used the multi-scale multi-angle convolution. For detecting the features, we are using the SURF method. It gives result at Octaves, Metric threshold and Scale level. These three features are combined to form ROI [i.e. Region of Interest]. Next stage is to apply hybrid DeepIRDI on extracted features. Using hybrid DeepIRDI we can easily retrieve classes.
Using inlier, we can recover scale and angle and estimate the geometric transformation.
Hyperspectral images have various features like Resolution, Light Intensity and angle. Pre-processing is applied on that image to obtain features. The Pre-processing includes removal of destruction noise. We have applied hybrid DeepIRDI algorithm on generated features which are further classified. The matching feature is generated and it contains two parts outlier and inliers. Here, we have done inlier selection because it gives more important and high priority features than the outlier.
Fig 1: Overview of proposed approach
Inlier selection M x N generates a probability matrix PR which is then taken by the transformation solver based on the Gaussian mixture model (GMM)[3]. The PR [m, n]
entry of this matrix is the probability of matching xn and ym. Assuming xn corresponds to ym, we obtain a large putative probability PR [m; n]. And a high probability would lead to a visible transition over ym which can be aligned with the respective pair of characteristics.
The probabilities are determined using both the convolutionary feature and the geometric structural information. The PR matrix is achieved by the following operation:
Prepare the convolutional cost matrix by
…... (1)
Eq1., shows that how to extract, match features of Hyperspectral images and detect objects in different location, different angle at different viewpoints. which is acquired by performing a χ 2 test. [1]. Both and are valued in [0, 1]. The following equation elaborate that, how the geometric transformation is estimated. Where (b) and
(b)denotes the no. of points decreased in the bth location point ym and xn respectively.[1]
… (2)
Using an element-wise Hadamard product, we calculate an integrated cost matrix C (denoted by ⊙):
Finally, the prior probability matrix is calculated using
… (4)
Then, we use the optimal parameters of transformation (W, σ^2, ω). The aim of such an approach is to maximize a likelihood function or minimize the negative log- likelihood function equivalently: Pold (m | xn) denotes a posterior likelihood. The equation may be rewritten after expanding this equation and omitting derivative redundant terms as:
… (5)
Algorithm: Multi Scale Multi Angle hybrid DeepIRDI Algorithm Input: IX and IY
1 Initialize parameters;
2 Pre-match and choose the convolutional feature point sets X and Y from IX and IY; 3Initialize ;
4do
5 For every k iterations:
6 Calculate the cost matrix for convolutionary features
7 as per to Equation 1.
8 Calculate the cost matrix for the geometric structure
9 as per Equation 2;
10 Calculate the cost matrix ;
11 Solve the cost matrix C linear assignment.
12 Calculate the prior probability matrix PR using Equation 4;
13 Update the threshold ;
14end 15E-Step:
16 Compute posterior probability matrix PO by 17
18end
19While there is no convergence of equation 5;
20 Calculate the transformed picture IZ using interpolation with thin plate spline.
21 Output: IZ
The remote sensing accepts input as hyperspectral image having various features like Resolution, Light Intensity and angle. Pre-processing is applied on that image to generate features. The extracted features from image are generated with the help of CNN method. On generated feature the Image Registration algorithm is applied. The matching
feature is generated and contains two parts outlier and inliers. By using inliers selection the object matching is done to provide accuracy.
3. Experiment Results
In our study, both qualitative and quantitative tests were used t o check the efficiency of CNN's multi-angle multi-scale model [3] of hybrid DeepIRDI [1]. This section focuses mainly on precision and other few parameters such as PSNR, Entropy, Kappa co. etc. on multi-label objects / images of satellite angle variants and the real SAR image datasets. We have split the table in few variations of angle, resolution and fusion of both angle and resolution. The result acquired is provided in table no I below. Based on the analysis of this table, we have discovered that PSNR remains in the range of 17 to 18 Db and that Entropy is in the range of 6.3 to 6.7 whereas SNR is 6.1 to 6.9 Db which does not hamper the image quality factor of the picture after all the preprocessing activities , which means we have successfully managed to get proper accuracy without modifying the image quality.
Also based on Angle θ it is observed that whatever rotations has been made in dataset creation i.e. Angle Variants are recovered accurately with minimal error of 0.1 to 0.6 degrees only with a negligible error. Hence, we have achieved angle invariant property as well in Image registration algorithm
Table I: Analysis of Angle, Resolution and Fusion of both.
Dataset S.
No
Variant Entro py
PSNR Kappa Co.
SNR Scale Recover
Angle θ Recover
Matchi ng%
Dataset A
1 6.46 17.97 1 6.12 1 0 98.65
2 6.47 17.81 0.66 6.28 1 30.14 88.10
3 6.47 17.85 1 6.25 1 180.20 86.56
4 27 6.48 17.75 1 6.34 1 270.04 77.25
Dataset B
5 64x64
6.30 17.23 0.79 6.86 1.68 0 85.10
6 256x256 6.46 17.94 0.97 6.16 1.7 0 80.39
7 512x512 6.46 17.84 1 6.25 1.70 0 76.19
Dataset C
11
6.30 17.95 0.75 6.14 1.71 90.2 80.5
12
6.46 17.87 0.99 6.22 2.69 89.90 73.83
13
6.46 17.85 1 6.25 3.70 90.02 85.71
14
6.60 17.57 0.82 6.52 1.76 180.26 84.76 15
6.46 17.94 1 6.15 2.70 180.09 83.09
16
6.46 17.83 1 6.27 3.70 179.98 74.23
17
6.30 17.61 0.9115 6.48 1.7375 269.67 90.81 18
6.46 17.91 0.6187 6.28 2.7051 270.145 82.92 19
6.64 17.87 1 6.22 3.6996 270.78 76.54
While analyzing accuracy factor we have found higher accuracy than any of the traditional method. As compared to our proposed algorithms it was 95% to 99%.
Previously it was 95% in Image Registration algorithms hence due to modified algorithms we have managed to improve accuracy by 4%. This comparison is shown in table II where we have tested the results with SIFT, Image Registration (IR) and our previous methodology MS-MA BT CNN [16] algorithms. Also, we have mentioned here that, our new algorithms, i.e. Multistage Multi angle DeepIRDI is showing higher accuracy of 98.6 % than our pervious proposed algorithm MS-MA BT CNN which has highest accuracy of 84% on same dataset [16].
Table II: Represents the comparison of hybrid DeepIRDI with the previous algorithm
Dataset hybrid
DeepIRDI (Ours)
MS-MABT CNN [16]
IR SIFT
Satellite 98.65 84.14 95.65 71.71
UAV 95.43 79.25 93.37 42.94
Table III: Represents the comparison of hybrid DeepIRDI with the previous algorithm with respect to plane dataset, angle variant, fusion and resolution variant.
In Table III we have represented detailed analysis of accuracy factor with respect to Dataset verses minimum, maximum and Average accuracy. We have found that average accuracy remains in the range of 80% to 87% whereas Max accuracy is 98.65% and minimum accuracy is 73.83% in case of fusion dataset. Also as compared to our pervious method MS-MABT CNN [16] all of the values i.e. avg, min, max were found to be higher in our proposed algorithm. Hence, we have successfully managed to improve accuracy of our proposed method by 12%
compared to that of our pervious method which was notifiable factor in this implementation.
Dataset Parameter hybrid DeepIRDI
(Ours)
MS-MABT CNN [16]
Satellite dataset with angle variant Dataset A
Avg 87.64 83.30
Min 77.25 82.57
Max 98.65 84.71
Satellite dataset with resolution variant Dataset B
Avg 80.54 82.75
Min 76.19 81.00
Max 85.10 84.14
Satellite dataset with Angle and resolution Fusion Dataset C
Avg 81.98 86.38
Min 73.83 84.49
Max 90.81 88.99
Graph 1: Entropy Results
We have represented graph 1 for entropy values of our dataset. In angle variant entropy was stable but in fusion it was randomly varying in -between 6.3 to 6.6. The higher value is noticed as 6.5 in fusion dataset of 270 degree and 512 x 512 resolution.
Graph 2: PSNR values Result
Above graph shows that all values of PSNR for angles, resolutions, and fusion datasets which are varying. Minimum value is 17.2 for 64x64 resolution and maximum value is 18. Average value is 17.6 for all dataset.
Graph 3: Kappa values result
The above graph represents the results of Kappa Coefficient values. In which maximum Kappa value is 1.09 and lowest value 0.6. In this graph all values are varying with respect to angle and resolution but its remain above 0.6.
Graph 4: SNR values result
The above graph representing the SNR value results. In which 6.8Db is maximum SNR value. Minimum value is in between 6 to 6.2Db. Average value is 6.5Db for fusion 180-64x64 dataset, which shows that, even if the images are distorted by angle or resolution, we have maintained noise level in all the data sets.
Graph 5: Scale Recovered values result
The graph 5 shows that 4 is maximum Scale value recovered. The values of angle are constant to 1 means there is no Scale variation since its only angle Variant dataset. The minimum value of fusion dataset is in between 1.5 to 2.
Average value is 2.5 for fusion 90x64, 180x256, 270x256 dataset. Therefore, we can demonstrate that whatever factor of scale was applied during the development of the dataset is effectively retrieved with minimal error.
Graph 6: Angle Recovered values result
The graph 6 shows, that angle recovered values lies between 0 to 270 degrees. As the rotation of dataset increases recovery value also increases. For resolution variant recovery value is zero since it hasn’t rotated. Above graph shows the angle invariant property of our algorithms.
Graph 7: Matching Percentage values result
The above graph represents, that 98.65% is maximum matching percentage value.
Whereas for older traditional algorithm like SIFT it is 71.71 % and for Image Registration (IR) it was 95.65 % hence our method is proven more accurate than existing algorithms. The average accuracy is 87.64 % which is also higher than the older algorithm.
Table IV: Comparison of GPU vs. CPU Execution Time
Algorithm MS-MABT CNN [16] hybrid DeepIRDI (ours) Time to execute
(CPU in Sec)
36 28
Time to execute (GPU in Sec)
17 11
We have tested and represented our system execution time on NVIDIA GPU Vs.
CPU and reported the result in the above table IV. Here we have found that, the GPU is giving faster result than CPU and which is further more better in performance as compared to our pervious algorithms MS MABT CNN [16] . We have managed to improve time by 6 Seconds in our proposed algorithms for all the mentioned datasets.
4. Conclusion
Our work describes the complete analysis with different algorithms of the Hyperspectral/ SAR image classification. Based on the above outcomes, it is clear that, our newly proposed hybrid DeepIRDI multi-scale multi-angle algorithm is effectively implemented and tested with 98.65 % accuracy rate with highest precision compared to other fundamental techniques and old algorithms. Also, our algorithm can be tested on real-time satellite datasets or any other hyperspectral images. We have already implemented our proposed algorithm & demonstrated on multiple combinations of datasets mentioned above. It is observed during the implementation that NVIDIA GPU played a very important role in reducing the complexity of time compared to the regular
CPU processor. In addition, the numerical indexes show that in most instances our method achieves good efficiency, which encourages the classification and interpretation of remote sensed images. In fusion with our existing work, we will implement a few more feature extraction techniques so, that it will be a comprehensive outcome of our research in the domain.
References
[1] Zhuoqian Yang 1, Tingting Dan2, And Yang Yang2,3 “Multi-Temporal Remote Sensing Image Registration Using Deep Convolutional Features”, IEEE Trans.,2018
[2] XiaomingTao, Chaoyi Han, Xiaowei Qin, and Jianhua Lu, “Multi-scale Convolutional Neural Network for SAR Image Semantic Segmentation Yiping Duan”, IEEE,2018
[3] L. G. Brown, ‘‘A survey of image registration techniques,’’ ACM Comput. Surv., vol. 24, no. 4, pp.
325– 376, Dec. 1992.
[4] K. Yang, A. Pan, Y. Yang, S. Zhang, S. H. Ong, and H. Tang, ‘‘Remote sensing image registration using multiple image features,’’ Remote Sens., vol. 9, 2017.
[5] Z. Wei et al., ‘‘A small UAV based multi-temporal image registration for dynamic agricultural terrace monitoring,’’, IEEE, vol. 9, 2017.
[6] J. Ma, J. Jiang, H. Zhou, J. Zhao, and X. Guo, ‘‘Guided locality preserving feature matching for remote sensing image registration,’’ IEEE Trans. Geosci. Remote Sens., to be published.
[7] J. Ma, Y. Ma, and C. Li, ‘‘Infrared and visible image fusion methods and applications: A survey,’’ Inf.
Fusion, vol. 45, pp. 153–178, Jan. 2019.
[8] B. Zitová and J. Flusser, ‘‘Image registration methods: A survey,’’ pp. 977–1000, Oct. 2003.
[9] Y. Ke and R. Sukthankar, ‘‘PCA-SIFT: A more distinctive representation for local image descriptors,’’
in Proc. IEEE Conf. Comput. Vis. Pattern Recognition., vol. 2, Jun./Jul. 2004, pp. 506–513.
[10] Wenping Ma , Jun Zhang, Yue Wu , Licheng Jiao, “A Novel Two-Step Registration Method for Remote Sensing Images Based on Deep and Local Features”, IEEE Transactions on Geoscience and Remote Sensing ,Volume: 57 , Issue: 7 , July 2019
[11] D. G. Lowe, ‘‘Object recognition from local scale-invariant features,’’ in Proc. 7th IEEE Int. Conf.
Comput. Vis., vol. 2, Sep. 1999, pp. 1150–1157.
[12] Y. Ke and R. Sukthankar, ‘‘PCA-SIFT: A more distinctive representation for local image descriptors,’’
in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., vol. 2, Jun./Jul. 2004, pp. 506–513. Conf. Intell.
Transp. Syst., Oct. 2013, pp. 1284–1291
[13] F. Dellinger, J. Delon, Y. Gousseau, J. Michel, and F. Tupin, ‘‘SAR-SIFT: A SIFT-like algorithm for SAR images,’’ IEEE Trans. Geosci. Remote Sens., vol. 53, pp. 453–466, Jan. 2015.
[14] A. Sedaghat and H. Ebadi, ‘‘Remote sensing image matching based on adaptive binning SIFT descriptor,’’ IEEE Trans. Geosci. Remote Sens., vol. 53, pp. 5283–5293, Oct. 2015.
[15] R. Jonker and A. Volgenant, “A shortest augmenting path algorithm for dense and sparse linear assignment problems”, Computing, vol. 38, pp. 325-340, Nov. 1987.
[16] Sujata Alegavi and Raghavendra Sedamkar, “Hyper Spectral Image Classification using Multi Labelled, Multi-Scale and Multi-Angle CNN with MS-MA BT Algorithm” International Journal of Innovative Technology and Exploring Engineering (IJITEE), Volume-8, Issue-9S3, July 2019.