Research On Pollen Image Recognition Algorithm Based On Multi-feature Fusion

The identification and classification of pollen is of great significance in biology and medicine,and it is widely used in many fields.The traditional pollen recognition mainly relies on the artificial microscope,which not only requires the operator to have relevant knowledge experience,but also the process of identification is time-consuming,and the recognition accuracy is low.With the rapid development of image processing and pattern recognition theory in recent years,the automatic recognition classification based on pollen image feature is an effective method to discriminate pollen image.Since the original pollen image was illuminated during the process of collection,noise pollution and other extermal factors,the quality of the pollen image has been affected by different degrees,resulting in a low recognition rate,so the characteristics of the extraction of pollen has a strong robustness,the extraction process has a certain real-time,and finally can obtain a higher recognition rate.This paper studies the fusion of multi feature extraction of pollen image,and starts from two aspects of feature descriptor and classification recognition algorithm,which not only enhances the accuracy of classification and recognition of the algorithm,but also improves the efficiency and real-time performance of the algorithm.The main contents of the study include:(1)Aiming at the traditional feature extraction algorithm of single pollen image,the distinguishing feature of extraction has the disadvantages of weak noise-resisting ability and low geometrical invariance,in this paper,the traditional machine-learning related image recognition theory is used to propose a new classification algorithm for pollen image fusion Zemike moment and Bof?surf feature fusion.Firstly,the zemike moment descriptor of the pollen image and the SURF feature descriptor based on the scale spatial gradient information are extracted,the K-means feature clustering of the SURF feature descriptor is constructed,the accelerated robust feature packet Bof-surf is built,and finally the two features are fused,Recognition classification is accomplished by SVM.Because the zermike moment and the surf descriptor are invariant,it has good robustness to the pollen scale and rotation,and the correct recognition rate is also high.(2)For the traditional pollen image classification algorithm in the process of extracting features,how to choose the most effective extraction features of a certain complexity,in this paper,using the depth machine learning image recognition framework,a feature fusion algorithm based on convolution neural network for the classification of pollen images is presented.This algorithm first standardized processing of pollen image,as the first training network input layer image data matrix,and then to the pollen image direction gradient histogram hog characteristic processing,as the second training network input layer image data matrix,the processing layer structure is a double-layer convolution-pool layer,The convolution matrix parameters of each training network layer are optimized and improved,in order to carry out more effective feature extraction for different data matrices,and then input the feature fusion layer added to this algorithm,then through the whole connection layer processing integration,finally through the output layer of Softmax and loss function to complete the model training,Finally used for classifying pollen images.