The Research On Automatic Identification System For The Microscopic Images Of No Chaetoceros Harmful Algae

In recent years, red tides with serious harm appear more and more frequently in coastal waters of China, which caused more attention by the governments, the public and scientists. The country has invested a lot on the researches on formation mech-anism, early warning, forecasting, preventing and controlling methods of red tide. Therefore, identifying the dominant species of red tide rapidly and effectively plays an important role in automatic monitoring of red tide. Based on the traditional biologi-cal morphological theory, this paper divides red tide algae in China's coastal waters into two categories(with chaetoceros and without chaetoceros), and designs the automatic identification system for the microscopic images of the 15 kinds of no chaetoceros harmful algae, by using the technology of the image processing and pattern recogni-tion. The work mainly contains:1. accomplishing the design of the automatic identification system for the micro-scopic images of the 15 kinds of no chaetoceros harmful algae, with the category features of each kind.2. proposing a method of image segmentation for cell image based on multiple directions projection and automatic thresholding. This method can extract the object cell from the microcopic image and has a good performance on images of the species without setae. Therefore, object cell information in microscopic im-ages can be extracted completely. Besides, the location of the cell in microscopic images can be located accurately to remove noises around the cell because this method adopts 8 directions projection.3. extracting the formfactors(7) and moment invariants(12) based on the biological morphological features of the no chaetoceros. And verifying their invariance for the trans formation of translation, scaling, rotation by experiments.4. constructing the identification model for the no chaetoceros harmful algae based on the SVM. And showing the effectiveness and feasibility of the features ex- tracted through a experiment, which contains 2400 images for training and 600 images for testing.