Researches On Automatic Recognition For Microscopic Image Of Harmful Algae Blooms

The21st century belongs to the ocean. While in recent years, marine disastersbecome the most serious impediment to the development and utilization of marineresources. The frequency of red tide which is one of the world wide marine disastersis becoming higher and higher in coastal waters of China. The red tides damage themarine environment severely, and impact on coastal marine economy and humanhealth. Therefore, our government has concentrated on early warning and preventionof red tides. And the key step is to recognize the dominant species which induce thered tides. According to the identified species, it can help us take some measures toprevent and control it.This thesis analyzes the bio-morphological of40kinds of harmful algae bloomswhich are common in the coastal of China. And these species have been divided intotwo categories (not belong to diatom and diatom). Then this paper builds amicroscopic images automatic diagnosis and recognition system based on imageanalysis technology. The main work of this paper is as follows:1. Species Classification. Divide the40kinds of species into two categories whichare diatom and others that not belong to diatom. Accomplish the design of theautomatic recognition system based on the idea of classification.2. Image Target Extraction. According to the characteristics of obscure edge of themicroscopic images, use an improved Otsu segmentation method to segment theimages. For the special morphological characteristics of diatoms, using themethod which is based on gray-scale model algorithm to segment the images ofdiatoms.3. Shape Feature Extraction. Based on the differences of shape characteristics of thevarious algae species, research the shape feature extraction. Extract19-dimensional feature vectors which include12moment invariant features and7shape features using the moment invariants and parameters of shape characteristic. Through the feature extraction for species, verify the effectiveness and invariancetransformation for the rotation, translation and scaling.4. Multi-class Classification. Research on SVM and trained the rough classificationmodel, diatom model, and the model for species which are not belong to diatom toidentify the species.Using the proposed method, the experiment have trained the models with4240images and verified the effectiveness with2006images. The average rate is82.7%,and the actual recognition rate is81.05%after removing the rough classification error.And it has verified the effectiveness of the method which is proposed by the paper forthe40species.