Identification and Enumeration of Waterfowl using Neural Network TechniquesMichael Cash12/19/03ECE 539: Introduction to Artificial Neural Networks and Fuzzy SystemsSemester ProjectProfessor Y. H. HuIdentification and Enumeration of Waterfowl using NeuralNetwork TechniquesMichael Cash12/19/03ECE 539: Introduction to Artificial Neural Networks and Fuzzy SystemsSemester ProjectProfessor Y. H. HuECE 539: Introduction to Artificial Neural Networks and Fuzzy SystemsMichael Cash Semester Project 12/19/2003 Background: The current status and past trends of waterfowl populations are annually studiedto estimate abundance, location, and species composition of ducks and geese throughout thecountry. Annual surveys provide most of the data for these studies, but the survey results arehighly controversial. Several techniques are used to count waterfowl; the two most accepted arethe area search (or direct count) method and the point count method. Area search involvesscanning a wide open area with binoculars and counting the number and type of waterfowl in thefield of view. Point count involves counting birds at a series of fixed, obstructed-view pointsthrough audio or visual tallying. Other methods, such as the Migratory Bird Harvest InformationProgram (HIP), rely on hunters to describe annual kills. Finally, some methods employ use ofairplanes or helicopters to extend the range of the area search.Most of these methods employ visual identification as the primary means of classification. Thisincludes large flocks of thousands of birds or more; however, in large flocks minority ducksbecome difficult to see and count. Additionally, ducks fly quickly and erratically makingpositive identification difficult. These and more sources of error add up quickly, and produce aresult that is not trusted by most hunters and some experts.Overview of Project: This project attempted to classify number and species of waterfowl fromdigital photographs. The premise was to replace the human eye as the sole classifier. Digital photographs from the web and from magazines were used as inputs. The first step in dataprocessing was scaling each pixel color to a useful range. Then, the number of ducks in thephoto were counted using a clustering technique. After each cluster (bird) is identified, a multi-layer perceptron was to be employed to try to classify the species and/or gender of the bird. Twodigital images photographs were used to set parameters and verify the functionality of thealgorithm, and 5 others were tested without modification. Data Processing: The first step was to pre-process the image by observing the file size andnumber of pixels in the image. As with all image processing algorithms, time and processorpower required increase dramatically with image width and height dimensions. Therefore, allimages were cropped to only include the areas of interest. The cropping operation provided2ECE 539: Introduction to Artificial Neural Networks and Fuzzy SystemsMichael Cash Semester Project 12/19/2003 another benefit: irregular backgrounds, such as trees or other vegetation, were easily eliminatedfrom the images. The final image ready to process consisted only of flocks of birds and somefairly consistent background, such as sky or water.All images were (re)formatted to be JPEG compatible. The JPEG images could be easilyhandled using MATLAB commands ‘imread’ and ‘image’, and were considerably smaller in sizethan alternatives such as bitmap (BMP). The format of JPEG is [v by h by 3] matrix: ‘v’ isvertical pixel dimension, ‘h’ is horizontal dimension, and the pixel content is stored in Red-Blue-Green (RBG) format in values of 0-256. The values are stored as 8-bit unsigned integers, whichrequired conversion to type double for analysis purposes.Background Elimination: An algorithm was developed to identify the foremost backgroundcolor, allowing separation between duck pixels and background pixels for use in the clusteringalgorithm (‘duck_background.m’). The nominal background color was taken as the median (notmean) pixel color of the entire image. This assumption required all graphics to contain at least50% background. The background colors of the images varied slightly with spatial pixel position because oflighting conditions, shadows, etc. Therefore, a range of pixel colors was declared to be‘background’ for processing purposes. This range was found by symmetrically adding a slightoffset to the nominal (median) background color. The most useful range eliminated most of thebackground but did not cause significant dropouts of pixels of interest (namely pixels located onbirds). This range, in 8-bit RBG format, was [nominal] ±[30 30 30].The coordinates of pixels of interest, or simply not background pixels, were placed in a matrix tobe clustered.Clustering: The matrix containing pixel coordinates of interest (non-background) was clusteredusing Prof. Hu’s ‘kmeansf.m’ k-means clustering algorithm. The two feature vectors were thehorizontal and vertical pixel coordinates. Before the algorithm could be called, however, cluster3ECE 539: Introduction to Artificial Neural Networks and Fuzzy SystemsMichael Cash Semester Project 12/19/2003 centers were initialized by assigning them to a random point around the mean pixel position. Itwas very important that the initial positions were random and were located across the full rangeof the image. If these two characteristics were not implemented, the algorithm would placemany cluster centers on top of each other. The number of initial cluster centers was held steady at 250 for all results in this paper. Thisvalue was found experimentally with two images: one image contained many (>50) birds, andthe other image contained ~15 birds..The k-means algorithm itself used the L2 norm and allowed for deletion of empty clusters uponconvergence. This was central to the idea of counting birds; if the number of cluster centers wasfixed it would be very difficult to use the k-means algorithm. The idea was to start with 250cluster centers and reduce the number down to the number of birds in the image.The fractional error between successive distortions for convergence was set to 1E-30, which wasalso found experimentally with two images. Maximum iteration number was set to 200, but wasnever
View Full Document
Unlocking...