Automate fry counting using computer vision and multi-class least squares support vector machine

In this paper, an approach based on geometric features to count overlapping fry fish is presented. Back propagation neural network (BPNN) and least squares support vector machine (LS-SVM) were used to construct classification models. 19 video clips with fish numbers varying from 10 to 100 were captured by a computer vision system. A total of 600 sub-images with overlapping fish were randomly selected, 300 images were used as a training set to create a calibration model, and remaining images were used to verify the model. 7 geometric features (area, perimeter, convex area, bounding box width, bounding box height, skeleton length, endpoint number) were obtained from the overlapping fish images. Results indicate that the best performance with about 98.73% of the average counting accuracy rate is achieved by LS-SVM model, which is better than the performance of BPNN model. The combined multiple geometric features coupled with an LS-SVM classifier is a highly accurate way for fry fish counting. (C) 2012 Published by Elsevier B.V.