Study Of Zooplankton Automatic Recognition Method For Dark Field Image

Ocean area accounts for about70%of the earth’s surface, and along with therapid development of marine economy, marine plays a very important role in theecological system of the earth. The development and utilization of marine resourceshas great impact on marine ecological system and environment. Marine zooplanktonbiomass, population structure, and community diversity, and life history play animportant role in marine ecosystems, marine biogeochemical cycles, marineenvironment, and global climate change research.Massive image data obtained by the optical imaging observation system in situpose a great challenge for marine plankton researchers, On the basis of surveying theexisting recognition methodology, this paper explored image preprocessing, featureextraction, classification and identification for dark field images, and the achievementhas established a frame work for dark field zooplankton automatic identificationsystem.For the image preprocessing phase, histogram equalization method is adopted toovercome concentrated gray value caused by the uneven brightness, and then themethod based on adaptive wavelet denoising is used to suppress image noise ofunderwater zooplankton, while retaining the original details of the zooplankton.Region of interest extraction using "area detector" method to extract interestedobjectivs from whole image. Sobel edge detection operator is used to detect the edgeof the target organisms and the watershed segmentation method is used to segmentzooplankton objective, our argument is that this algorithm can split to some extenttouch or overlap objects and avoid contour offset and zooplankton image noise causedby the over-segmentation phenomenon.For the basic features selection, HU’s invariant moments, Fourier descriptors,gray level co-occurrence matrix(GLCM)texture and local binary pattern texturefeatures are evaluated by a single support vector machine classifier. The results showthat combined HU’s moment invariant features with GLCM texture featurescombination got83.11%recognition rate and the local binary pattern texture featurecombination achieved a83.2%recognition rate respectively.Finally a dual classifier scheme is proposed and each layer consists of a supportvector machine. In the first layer the basic morphological features are as input to getclassification category, and in the second layer the global or local texture features areused as input to get another category, if these two categories are the same, the samplefor this category is true, otherwise it will be classified as a class which cannot berecognized. The ROC curve is used to evaluate the classifier performance. Theclassification results show that dual classifier scheme has better performance than asingle classifier. With the training samples increased, the performance with local binary pattern feature has been improved. And the effect of local binary pattern isbetter than gray level co-occurrence matrix features.