Research On Robust Feature Extraction Algorithms For Pollen Images

Effective recognition and prediction of pollen type is of great practical value in the fields of plant selection breeding, pollen allergy prevention and treatment,paleoclimatic reconstruction, etc. Traditional pollen recognition techniques mainly rely on microscope examination, requiring great expertise in palynology. The process is also time consuming and laborious,and the recognition result is susceptible to the subjective influence of the experimenter, which may result in poor recognition rate.Due to the fact that the microscopic pollen images have similar texture and structural features to the ordinary images, and the characteristics of different types of pollen images are highly distinguishable,the automatic extraction and classification of pollen image features based on pattern recognition theory has become the mainstream approach for pollen recognition, which has been extensively studied.However, problems such as low recognition precision and poor real-time performance affected by factors such as illumination, noise and contamination are restricting its practical application.As a result, improving the robustness and real-time response of the recognition system is the key to make the pollen automatic identification technology more practical. Therefore, this paper studies the versatile feature extraction method of pollen images, in order to improve the robustness of features which are invariant to geometric transform, such as image rotation and scaling, and improve the classification accuracy and efficiency. The contents of research include:(1)In view of a single scale space is not capable for characterizing the detail features of all the pollen images, a multi-scale subspace covariance descriptor based on tower wavelet transform is proposed to improve the robustness of features via texture and multi-scale fusion. First, the different sub-band spaces of the image are constructed by using the tower wavelet transform. By further calculating the characteristic parameters of the gray level co-occurrence matrix, the sub-band co-moment descriptor of the multi-scale space (MSSCM) is extracted. The experimental results on Confocal and Pollenmonitor image databases show that the descriptor can describe the texture distribution of pollen images effectively, which can speed up the classification process and has good robustness to the rotation and pose change of pollen images.(2)In order to solve the problem of high feature dimension, rotation and significant change in configuration of pollen images, a new descriptor based on the Gaussian differential pyramid of partial binary roundness mixture texture is proposed(MSLBPC). The method first divides the image into the upper level of the texture and the prominent upper level of the shape in the Gaussian difference pyramid. The feature points of the texture level are then extracted by using the shape-based SIFT algorithm, and the texture descriptor is combined with the LBP to reduce the feature dimension number. The shape descriptor is further generated through the shape prominent level based on the roundness of geometric features to reduce the redundant information and restrain noise disturbance. The final feature combined the aforementioned two types of descriptors using weighted linear combination. The experimental results on two datasets show that the proposed descriptor can describe the image well and has good robustness to the rotation, translation, scaling and noise disturbance of the image.