Research On Feature Extraction Algorithms For Pollen Images

The automatic classification and identification of pollen images play an important role in environment, medicine and other fields. But at present, the traditional pollen classification and recognition mainly relies on artificial, which usually needs professionals have enough knowledge of pollen morphology and practical experience. This work is also very time-consuming and the recognition accuracy will inevitably be influenced by personal subjective factors, which results in the low identification efficiency. In recent years, with the development of image processing and pattern recognition theory, using computer for automatic classification and recognition of pollen images has become an effective means to solve these problems. In order to improve the robustness of features to image geometric transformation and the accuracy of classification, research on feature extraction algorithms for pollen images is performed in this article. The main contents of this article include:(1) According to the problem that the existing texture descriptors are mostly dependent on the average grey value, which is easy to cause the loss of image information, a new roughness descriptor based on Gaussian scale space is proposed for pollen image classification and recognition. First, the Gaussian convolution is used to build the Gaussian scale pyramid of the image. Second, the statistical distribution of roughness frequency is extracted to build the Scale-Space Roughness Histogram Descriptor (SSRHD). The experimental results on Confocal Dataset and Pollenmonitor Dataset demonstrate that the SSRHD can effectively describe the pollen image texture and has good robustness to pollen rotation and pose variation.(2) According to the problem that the redundant information of existing shape descriptors is so much and the characteristic dimension is high while the recognition rate is low, a new Fourier descriptor based on optimized 3D Zernike moments(ZMFD) for automatic pollen images recognition is proposed. First,3D Zernike moments of the image are extracted in spherical coordinates. Second, the genetic algorithm based on probability is applied to filter the Zernike moments in order to reduce redundant information. At last, the Fourier transform coefficients of optimized Zernike moments are computed and the normalized Fourier descriptor is extracted as the last feature descriptor. The experimental results on Confocal Dataset and Pollenmonitor Dataset show that the descriptor can effectively describe the pollen image and is robust to the rotation, translation and scaling of the image.(3) The comparative analysis experiments of the texture descriptor and the shape descriptor. Comparative analysis is performed between the two descriptors. There comes to the results that different types of pollen images have different shape and texture characteristics, and the above two methods have different adaptability to different types of pollen images.