This paper presents a novel rotation-invariant texture image retrieval using particle swarm optimization (PSO) and support vector regression (SVR), which is called the RTIRPS method. It respectively employs log-polar mapping (LPM) combined with fast Fourier transformation (FFT), Gabor filter, and Zernike moment to extract three kinds of rotation-invariant features from gray-level images. Subsequently, the PSO algorithm is utilized to optimize the RTIRPS method. Experimental results demonstrate that the RTIRPS method can achieve satisfying results and outperform the existing well-known rotation-invariant image retrieval methods under considerations here. Also, in order to reduce calculation complexity for image feature matching, the RTIRPS method employs the SVR to construct an efficient scheme for the image retrieval.
With the advancement in information technology, people can readily take pictures using digital cameras, camcorders, web camera videos and smart mobile phones anytime and anywhere. This phenomenon makes image databases get an explosive growth. Thus, users need to employ efficient and effective tools to search for the images they want in the huge image database [7]. Current popular websites, such as Google and Yahoo, provide the function for image searching based on the text annotations of images. As a result, users often spend a lot of time and effort to edit these text annotations for each image. However, in many cases, text annotations cannot clearly describe the image contents. Moreover, a query request is difficult to be precisely described by text annotations. Therefore, the content-based image retrieval (CBIR) technique has been proposed to overcome these limitations mentioned above. The index in every image in the CBIR is popularly formed by the composition of image's own visual contents [16]. Subsequently, researchers improve the CBIR with rotation-invariant. At present, rotation-invariant texture image retrieval becomes important issue in image retrieval.
The texture descriptor of images plays an important role in computer vision and image pattern recognition, especially in representing the contents of an image [4] and [18]. Several image retrieval systems, which extract the rotation-invariant texture features of images, have been developed recently [5], [7], [8], [11] and [19]. The feature extractions of several proposed methods have been devised in frequency domain for extracting rotation-invariant features [5], [7], [11] and [16]. In Ref. [7], Kokare combines a dual-tree rotated complex wavelet filter (DT-RCWF) and a dual-tree complex wavelet transform (DT-CWT) to obtain the texture features for rotation-invariant from 12 different angles. However, the similarity measurement formula is not optimized. Thus, the method cannot get better discrimination between two different images [7]. In [16], Tzagkarakis presents the kullback-leibler-divergence (KLD) method which employs the Gaussianized steerable pyramids to extract the texture features of images. Nevertheless, it is insufficient to search for the optimal number of outputted images. In Ref. [5], the rotation-invariant Gabor (RIG) method is proposed which combines the Gabor filters with same scales and different angles to extract the rotation-invariant texture features of images [5]. In Ref. [11], Rallabandi presents wavelet-based hidden Markov trees (WBHMT) which combines the wavelet transformation and the hidden Markov tree to extract the rotation-invariant texture features of images. Unfortunately, the feature extraction algorithm of the method requires high computational complexity. In Ref. [12], Sim presents the modified Zernike moments (MZM) which combine the discrete Fourier transformation and the Zernike moments to construct the rotation-invariant texture features of images. Nevertheless, the similarity measurement formula is not optimized. Hence, the method cannot get nearly optimal solutions in the number of outputted images, various feature weights, etc.
Aforementioned phenomena motivate us to develop a novel rotation-invariant texture image retrieval method, called the RTIRPS method which can overcome the drawbacks of the above methods [5], [7], [11], [12] and [16]. The RTIRPS method employs the PSO algorithm which searches for a set of nearly optimal thresholds and enlarging constants in order to improve the retrieval performance. Fig. 1 depicts the conceptual design for the RTIRPS method. Note that the RTIRPS method has following properties: (1) simultaneously employing three rotation-invariant features to design dis-similarity measurement algorithm, (2) optimizing the dis-similarity measurement algorithm by using the PSO algorithm, (3) searching for an optimal number of outputted images, (4) reducing the computational complexity of feature matching by employing the SVR.
The remainder of this paper is arranged as follows. Section 2 introduces the background knowledge including the PSO algorithm, the SVR, the LPM, the Gabor filter, and the Zernike moment. Subsequently, the RTIRPS method is described in Section 3. Section 4 presents the experimental results. Finally, conclusions are given in Section 5.
This paper has proposed the RTIRPS method which can be employed to retrieve rotation-invariant gray-level texture images. First, it respectively utilizes the LPM combined with the FFT, the Gabor filter, and the Zernike moment to extract three kinds of rotation-invariant features for the gray-level texture images. These three feature classes are subtly integrated to form a combined feature vector. The main contribution of this paper is to develop a novel procedure which calculates feature difference for each individual feature component. Subsequently, the PSO algorithm is utilized to optimize the RTIRPS method via finding out a nearly optimal parameter set, ρ0, which is employed to character the RTIRPS method. It is utilized to optimize the RTIRPS method for these parameters while using the measurement procedure of dis-similarity degrees for two different images. Finally, experimental results demonstrate that the RTIRPS method outperforms other existing well-known rotation-invariant image retrieval methods under considerations here. Furthermore, in order to reduce computational complexity for image feature matching, the RTIRPS method also utilizes the SVR to construct an efficient scheme for the image retrieval.
