Microscopic Image Segmentation And Feature Extraction Of Harmful Algae Based On Biological Morphology

In recent years, more and more frequent HABs(harmful algae blooms) have posed serious threats on coastal environment, marine resources and public health. It is a grievous global marine disaster causing billions of economic loss every year in China. Governments and the scientific community are concerned with this situation and pay much attention to this issue. Artificial observation and analysis cannot satisfy the prediction of red tide due to the limitation of their efforts and biological knowledge. It is thus very urgent to study the methods of warning, forecasting red tide and establish operational monitoring system. It should be noted that identifying the dominant species of red tide plays an important role in automatic monitoring of red tide.On the sum-up of 40 algae species in the coastal waters of China Sea, and the analysis of their biological characteristics, the following conclusions are gotten:shape, cingulum, spine, seta, etc. are the dominant biological characteristics which can be utilized to recognize these species. Taking these biological characters into account, the micro-images recognition frame is designed.1. Image segmentation is a necessary pretreatment step in many target recognition applications. A coarse classification scheme is presented according to several characteristics of micro-images such as noises and poor constrast and so on.1) Minimum error thresholding is proved to be suitable for algae without seta by theorization and experiments. Aiming at the defect of thresholding methods, a target extraction method based on thresholding and projection intergal on multiple directions is proposed. The noises are removed by searching for the largest integral and seting other lesser integral zero. The results show that the approach proposed can extract cell object exactly.2) According to the specific problem of micro-image of Chaetoceros, the orientation angle model of gray image is established, in which seta components are reserved by decomposing the orientation angle vector into two gray images on X and Y axises. Then combining the filtering and morphological operations, Chaetoceros image segmentation is realized. The results show that the method outperfom thresholding as majority of seta components are extracted.2. This thesis focuses on biological features extraction based on biological characters of algae without seta. And then feature set is selected and established by combining the universal with domain-correlative characters.1) A method for extracting spine feature based on mathematical morphology is proposed, in which the pixel-width is introduced and the optimal structure element is selected automaticly by pixel-width histogram and area distribution.2) To tackle the over-segmentation problem in watershed algorithm, a method based on dominant gray levels and constraint marker watershed is proposed to extract cingulum region. In this algorithm, original image is reconstructed by dominant gray levels, eliminating local minima and noise disturber. Then markers of reginal minima are extracted from gradient image by using threshold, and imposed by shape, area and centroid constraints. The watershed transformation of the maker-modified gradient image is performed to achieve the cingulum feature extraction.3) Besides, after global shape features, texture features and domain specific features are analysised and extracted, C-SVM classifer is designed by "one-against-one" approach, recognition results of 15 species of algae without seta are received.3. This thesis deals with the research on Chaetoceros object representation and feature extraction based on skeleton. The skeleton tree is used to represent the Chaetoceros and geometrical features are extracted from it based on skeleton theory.1) A skeleton hierarchical decomposition approach based on competitive mechanism is advanced, in which the competitive mechanism is defined as the branch with the consistent direction with rachis element succeeds to trace, guaranteeing the rachis element integrity. The Chaetoceros object is represented by attributed tree by forming the rachis elements and branches into it, and the hierarchy of the tree and the connection relations of the nodes reflect the topological characteristics of skeleton.2) Considering the geometric characters of Chaetoceros object, combining the topological difference and geometric difference as the feature set, the similarity measurement suitable for micro-images of Chaetoceros is established.