Automatic In Situ Identification of Plankton M. Blaschko, G. Holness, M. Mattar, D. Lisin, P. Utgoff, A. Hanson, H. Schultz, E. Riseman, M. Seraki, W. Balch, B. Tupper Problem: Motivation: • Identify taxa of phytoplankton from images taken in situ by Flow. CAM • Phytoplankton is the basis of the food chain for marine life • Integral component of global carbon cycle • Studying abundance of different species important for understanding of global and local ecology System Overview Flow Cam • Manual identification is a daunting task, so automated solution is needed Raw Images Feature Extraction Ground Truth Labeling Feature Vectors Feature Selection Classification Ensemble I 1 f 12 f 11 C 1 I 2 MATLAB and C/C++ f 21 f 22 C 2 Vote Class Label I 3 f 31 f 32 f 31 C 3 Image Acquisition Challenges • Flow. CAM produces thousands of images in short time • Low magnification to increase field of view, resulting in low resolution images FFT • Images contain any type of organism, i. e. not restricted to any particular taxon • Flow. CAM developed at Bigelow Labs • Power Spectra of ADIAC 100 x images (top) and Flow. CAM 4 x images (bottom) FFT • Water siphoned directly from the ocean • Particles exhibiting florescence are imaged Features Feature Space • Cells categorized visually by shape and texture Shape Features Texture Features • • (h 2, w 2) h 2 • Perimeter • Gaussian Differential • Area • Co-occurrence • Moments Coordinate system Each point represents an instance 2 D Example using height and width Height and width are features • Local point features (h 1, w 1) h 1 • Convexity • Contour statistics w 2 w 1 • Contour spectrum Classification T classifier inducer • • Instance: x = <x 11, …, x 1 d> 1 Class label: Yi є { c 1, …, c. K} class labels Labeled instance: (xi, yi) Training set: T= {(x 1, y 1), …, (x. N, y. N)} Partition feature space into regions Each region contains instances in a class Classifier Induction: Estimate function mapping instances to class labels • Sometimes estimates commit errors Single classifier Results Classification Methods • K-Nearest Neighbors Ensembles T • Decision Trees • Naïve Bayes T • Combined estimates can lead to increased accuracy CI 1 CI 2 Vote • Ridge Regression • Support Vector Machines T • Methods used: Boosting, Bagging, and Multi-Classifier CIM Ensemble classifier Results • Improvements possible if individual classifiers are independent Conclusion • Combinations of shape and texture performed best Experiments • 980 expert labeled Flow. CAM image • pool of 780 total features • 10 -Fold Cross Validation • Best results with Support Vector Machines Best accuracy was 73%, comparable to consistency rate of human experts Future Work • Automated Feature Selection • Improved ensemble performance gains by inducing classifier independence • Experiments with Local image Features
Скачать презентацию Automatic In Situ Identification of Plankton M Blaschko
c436dc30fa60dbde93db14002030f660.ppt