Research On Texture Feature Extraction Method Of Pollen Images

The classification and identification of pollen particles has a wide range of applications in pollen allergen control,criminal investigation,oil exploration,paleoclimatic reconstruction and some other fields.The traditional identification of pollen grains is mainly done by artificial inspection under microscopy,which requires the operator to have a rich knowledge of pollen morphology.The identification process is time consuming and laborious,and the identification result is susceptible to subjective opinions,resulting in a relatively low accuracy.Since pollen images under microscope have similar contours,textures and structural features as common images,it is an effective way to extract pollen features and classify them automatically by computer.However,the existing methods for the classification and recognition of pollen images still have some shortcomings,including the following two aspects: Most of the existing descriptors are sensitive to noise and have poor robustness to the rotation and scaling of pollen images;Most descriptors fuse multiple features so as to construct the optimal respresentation of pollen images by taking advantages of different features,which greatly increases the time complexity of the algorithm and not conducive to the actual classification of pollen images.In view of the above problems,this thesis studies the texture feature extraction method of pollen images.The main contents of this thesis include:(1)According to the problem that the traditional Local Binary Pattern(LBP)has the shortcomings of noise sensitivity,low robustness to the rotation of image,etc,we improved Local Binary Pattern and proposed the Dominant Gradient encoding based Local Binary Pattern(DGLBP)descriptor,which has been applied to the identification of pollen images.Firstly,the gradient magnitude of the image block in the dominant gradient direction was calculated.Secondly,the radial,angular and multiple gradient differences of the image block were calculated separately.Then,binary coding was performed according to the gradient difference of each image block.The binary code was assigned weights adaptively with reference to the texture distribution of each local region,and the texture feature histogram of pollen images in three directions was extracted.Finally,histograms of texture features under different scales were fused,and the fusion features were used for the classification and recognition of pollen images.The experimental results on Confocal and Pollnmonitor datasets show that the proposed method is robust to the noise,rotation and scaling of pollen images.(2)According to the problem that the variation range of the texture of pollen images is smaller than that of ordinary texture images,and the wide quantization interval can not capture the subtle texture difference between various pollen images,a Local Decimal Pattern(LDP)was proposed,which has been applied to the identification of pollen images.The method captures the subtle texture difference of pollen images by increasing the number of quantization intervals and narrowing the range of the quantization intervals.First,the pollen images' gradient was decomposed into 8 directions to find the maximum,minimum and median gradient directions.Secondly,calculating the gradient amplitude of pixel blocks in each gradient direction.Then,decimal coding is performed according to the number of pixel blocks in each quantization interval.Finally,the LDP feature histograms in the maximum,minimum,and median gradient directions were calculated,and the fused features were used for the classification and recognition of pollen images.The experimental results show that the average correct recognition rate of the proposed method can reach more than 90%,and the recognition efficiency is also faster than some of the contrast method.