Research And Application Of Illuminant Invariant Features In Image Retrieval And Recognition

After several decades of development, technologies of image retrieval and recognition have made great achievements and been widely applied in many fields, such as industrial manufacturing, finance, public security, judicial branch and military. They are not only fast and highly efficient, but also have superior accuracies compared with human. There is a close relationship between image retrieval and recognition, the achievements of image retrieval technology further promote the advance of image recognition, which has been reflected in the application of handwriting recognition particularly.Illumination condition is one of the main challenges in image retrieval and recognition. Because light changes the distribution of image in gray space, affects the accuracy of retrieval, limits the advanced promotion and application of image recognition system. Based on the theory of multiscale geometric analysis, the main research of this paper focuses on extraction and application of illuminant invariant feature, which will further improve the accuracies of image retrieval and recognition.The main work is as follows:①Considering the lack of time-shift invariance and directional information in discrete wavelet transform and the aliasing problem in Contourlet domain, we proposed a novel image retrieval algorithm of extracting illuminant invariant feature by combing non-aliasing Contourlet transform (NACT) and denoising model. First, the image is converted to the logarithmic domain, and then we extract illuminant invariant feature with the method of NACT and denoising model. Finally retrieval is completed by using K-nearest feature line classifier. Experimental results show that the recognition rate of improved algorithm is higher than that of discrete wavelet transform combined with denoising model by 5.48%, compared with methods based on morphological features texture increased 13.38%.②According to the illumination invariance of phase information of the pyramidal dual-tree directional filter bank (PDTDFB) transform, an image retrieval algorithm of extracting illuminant invariant feature is proposed. First, we obtain a robust illumination feature of image using PDTDFB and relative phase depicted by Vonn distribution model. Then, the resistor-average(RA) distance is used to measure the similarity between images using the extracted features. Our experimental results show the effectiveness of the proposed algorithm. Compared with algorithm based on moment invariants, the recall rates improve by 1.83%-13.42%, precision rates increase by 10.76%-21.22%.③An illuminant invariant feature based on relative phase information is applied in recognition of Chinese handwriting image. Although the light changes of handwriting image are not that obvious, relative phase information with illuminant invariance could further improve the performance of recognition systems. First, we obtain the feature description of image using PDTDFB, magnitude described by derived magnitude distribution (DMD) and relative phase which is depicted by Von Mises(VM) distribution and Wrapped Cauchy(WC) distribution model. Then we apply the RA distance to measure the similarity between images using the extracted features. Finally, motivated by the idea of multiple attribute decision making (MADM), the two kinds of features extracted from magnitude and relative phase are fused appropriately to realize writer identification. Experimental results show that our method performs better than other traditional methods, the accuracy reaches 100% in the best situation.