SAR image classification method based on Gabor feature and K-NN

: Synthetic aperture radar (SAR) image target classification is a hot issue in remote-sensing image application. Fast and accurate target classification is important in both military and civilian fields. Consequently, this study proposes a novel SAR image target classification method based on Gabor feature extraction and K-NN classifier. First, the multi-scale Gabor features of SAR image are extracted. Then, a k-nearest neighbour (k-NN) classifier with principle component analysis is trained by the extracted Gabor features. Finally, the classifier is used to realise the multi-types SAR image targets classification. MSTAR database is used to validate the classification ability. Experimental results demonstrate that the proposed method has superior performance in term of efficiency and accuracy.


Introduction
Synthetic aperture radar (SAR) image target classification and recognition is of importance in both military and civilian fields. Fast and accurate multi-targets classification is of great help to many tasks, such as marine surveillance, fisheries management, and emergency response.
Researches on SAR image target classification are numerous, and several approaches are proposed in recent years. Mostly methods are based on traditional machine learning, Su et al. [1] proposed a SAR image classification method based on conditional random fields with multi-scale region connection calculus model. In [2], Karine et al. exploited a method combined saliency attention with SIFT keypoints for the SAR image target automatic recognition. The polar scale-invariant feature transform (PSIFT) is proposed in [3], which abandons the calculation of dominant orientations, and the targets are classified by a new matching method which is called the SAR-DM. Feng et al. [4] proposed a bag-of words (BOV) based on clonal selection algorithm for SAR image classification. These algorithms can realise the classification tasks to certain extent, while the accuracy is not satisfied. Recently, with the rapid development of artificial intelligence, the methods based on deep learning are proposed for SAR image target classification. In [5], David et al. proposed a method for SAR ATR by training convolutional neural networks (CNN). Ali et al. [6] proposed a method based on CNN with convolutional autoencoder. Chen et al. [7] proposed a method based on CNN with sparse auto-encoder. While these methods based on deep learning can achieve a high accuracy but are of low efficiency.
To achieve both high accuracy and efficiency, we propose a novel SAR image target classification method based on Gabor feature extraction. As shown in Fig. 1, the proposed method can be divided into three steps. First, the Gabor features of SAR image are extracted from seven scales and ten orientations. Second, the features are selected by principle component analysis (PCA) and a k-NN classifier is trained by the selected features. Finally, the SAR images are input into the classifier to realise targets classification. Our method is tested on the MSTAR database to demonstrate its outstanding performance on both classification accuracy and efficiency.

Gabor feature extraction
Selecting a suitable method to extract the target feature is a first and critical step to achieve fast and accurate target classification. Many feature extraction methods are proposed for target classification, such as scale-invariant feature transform (SIFT), the bag-of-visual words (BOV), grey-level co-occurrence matrix (GLCM), Gaussian Markov random field (GMRF) and so on. Therefore, Gabor wavelets, which have a good performance in extracting the target's local spatial and frequency domain information, are widely used in visual information explanation. Compared with other methods, Gabor wavelets transform has lots of advantages on feature extraction. First, the amount of data processed is fewer than other feature extraction methods. Consequently, it can meet the requirement of real-time processing. Second, the Gabor wavelets transform can deal with a certain degree of image rotation and deformation, so that the system robustness can be improved. Moreover, Gabor wavelets are sensitive to the edge feature of SAR image. In summary, Gabor wavelets can capture the properties of spatial localisation, orientation selectivity, and spatial frequency selectivity.
To extract better image features, the Gabor kernels are set at seven scales and ten orientations. Essentially, the Gabor feature The size of images processed is 128 × 128 pixels. When the features are extracted directly, the dimension of feature will be pretty high. As a result, it is not conducive to the subsequent classification process. To deal with the problem, the images are divided into sub-graphs with the size of 31 × 31. Then, the energy of each sub-graph is calculated. Finally, the energy matrices of all the sub-graphs are reduced in dimension and combined to a row vector. The row vector will be used as the Gabor feature of the image at a certain scale and orientation.

Principle component analysis (PCA)
When the number of the samples is known, there exists a maximum threshold of the number of features. If the number of actually used features exceeds this maximum, the performance of the classifier is not improved but degraded. This phenomenon is referred as dimensionality disaster in pattern recognition. Although the process of Gabor feature extraction can reduce the dimension to a certain degree, the dimension of the extracted features is still too high to cause the dimensionality disaster. To solve this problem, the principal component analysis (PCA) is applied. The essence of PCA is to map the sample data in high-dimensional space to lowdimensional space by linear transformation under the premise of representing the original data as best as possible. After linear transformation of the extracted Gabor features, the better distinguishable features in the original high-dimensional space are retained by larger weights, while other features lacking distinction are given lower weights. In our method, the 80% compression ratio of PCA is applied to the extracted Gabor features whose dimension is 7700. Then, the dimension of the features can be reduced to 1540. As a result, redundant features are removed and the performance of classifier can reach to the best.

Trained the cosine K-NN classifier
Since the KNN mainly depends on the neighbouring finite samples to determine the classifier, rather than by classifying the class fields. Therefore, the KNN method is more suitable than other methods for the sample sets with more overlapping or cross-class fields, such as MSTAR dataset.
Consequently, K-NN is chosen as the classifier in our method. K-NN classified by measuring the distance between different eigenvalues. The classifier is trained by the selected features in K-NN model. The common classification distance includes Euclidean distance, Minkowski distance, and cosine similarity. Compared to other distance calculation methods, the cosine similarity mainly distinguishes the difference from the orientation, but insensitive to the absolute value. From the perspective of SAR image data diversity, cosine similarity is the most suitable distance calculation method. Consequently, the cosine similarity is selected in our method. Cosine similarity uses the cosine value of the angle between two vectors as the measurement of the difference between two samples. The cosine similarity can be defined as follows: where A and B are row vectors, n is the dimension of the vector.

Experimental results
To evaluate the performance of the aforementioned SAR classification method, we use MSTAR database in our experiments. MSTAR is main about three classes tank targets whose images were captured under different depression angles coverings fully 0 to 360 range. The basic information of the database is shown in Table 1. The optical and SAR image of three types of tanks are illustrated in Fig. 3. Then, the last three rows are, respectively, shown the tank located in 0°, 45°and 135°.

Parameter finetune
Through the experiment, the Gabor features at 10 orientations has the best performance. Based on this result, the Gabor features at different scales are extracted. The recognition result using the extracted features is shown in Table 2. The experimental results demonstrate that the Gabor feature at seven scales is of the best classification capability. In the proposed classification method, the K-NN model is used to train the classifier. The number of neighbours is set to 10 by cross-validation and different distance calculation methods are applied to train the classifiers. The classification results are shown in Table 3. The cosine similarity is the most suitable distance calculation method.
In order to avoid the dimensionality disaster and improve the processing efficiency, PCA is used in the proposed method. The classification result in different compression ratio of PCA is shown in Table 4. The classifier is trained by cosine K-NN model. When the compression ratio is 80%, the extracted features have the best distinguish ability.

Comparison and results
To further demonstrate the outstanding performance of the proposed method, the experimental results are compared with other existed classification methods. In [7], Tao proposed a SAR automatic target classification method based on slow feature analysis (SFA). Cui proposed a hierarchical propelled fusion strategy for SAR target classification in [8]. Besides, Zhou proposed an efficient SAR ATR algorithm based on the contour and shape context in [9]. In addition, Ma proposed a coherent point drift-based scattering centre matching method to realise the ATR for SAR image in [10]. They all chose samples from BMP2, BTR70, and T72 for training as above. The classification results of [7][8][9][10] were given in Table 5 for comparison. From above experiments results, the method based on Gabor feature extraction does achieve a good performance. The classification accuracy rates of three classes reached to 99.15, 98.47, and 100%, respectively. The all-over classification accuracy is 99.41%.

Conclusion
The paper proposes a novel SAR image target classification method based on Gabor feature extraction. Since the Gabor features have a good performance in extracting the edge texture of SAR target and the KNN is more suitable than other methods for the sample sets with more overlapping class fields. In the proposed method, the Gabor features are extracted at seven scales and ten orientations. Then, PCA is used with 80% compression ratio to remove the redundant features. Finally, the cosine K-NN classifier is trained by the selected features. Experimental results demonstrate that our method can classify the SAR image target in high accuracy and efficiency.