Research On Local Feature Extraction Algorithms For 3D Pollen Images

The automatic classification and recognition of pollen play an important role in environment,immunology and many other fields.At present,the traditional pollen classification and recognition mainly relies on manual operation,which usually need operators with enough knowledge of pollen grains and the results are usually depend on personal experience.This work is also very time-consuming which results in the low identification efficiency.With the development of computer vision theory,using computer instead of humans for automatic classification and recognition of pollen images has become an effective way to solve these problems.The current researches on automatic pollen recognition are mainly focus on 2D pollen images.However,with the development of confocal microscopy,the quality of 3D pollen image has gradually increased,which lays a foundation for analysis of pollens' internal structure.The main contents of this paper include:(1)According the problem that two-dimensional feature cannot describe the internal structure and three-dimensional spatial pixel correlation of pollen image,this paper presents a local binary pattern feature for three-dimensional pollen images recognition.In this method,the feature plane is selected to mark the changing direction of local gray scale,and then the local gray scale vector on the center pixel neighborhood is calculated for constructing the optimal feature plane according to the local gray vector.The local texture feature on the optimal feature plane is extracted to construct the feature matrix.The statistical histogram descriptor of the matrix is finally extracted as the discriminant feature for the three-dimensional pollen image recognition.Experiments are performed on Confocal and Pollenmonitor,two standard European pollen databases.The results demonstrate that the best recognition rate of the algorithm can reach over 95%.Compared with the traditional algorithms,it has better robustness to the scale and attitude change of the pollen images.(2)According to the problem of excessive redundant information,high feature dimensions and low recognition rate in existing 3D local feature descriptors,this paper presents a 3D local key point extraction method based on SIFT for automatic pollen classification.In this method,3D Gaussian pyramid is constructed to obtain muti-scale pollen images,and then compute the local differential vector to explore local key points,then filter the key points by inter-layer contrast.The statistical histogram descriptor of the key points is finally extracted as discriminant feature of automatic classification of 3D pollen images.The experimental results on Confocal Dataset and Pollernnonitor 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 3D local features is presented in this paper.Comparative analysis is performed between the two descriptors.There comes to the results that different types of pollen images have dififerent shape and texture characteristics,and the above two methods have different adaptability to different types of pollen images.