Hierarchical classification, countingand length measurement of fishusing a stacking model approach

In this paper, the development of a hierarchical fish
classification framework is presented. The conventional data
collection technique for the commercial fish stock assessment
is a labour intensive and time consuming procedure. The purpose
of this project is to develop a framework that classifies
fish species on two level semantic hierarchy label, to count the
number of fishes and to measure the length of four different
fish species using a small dataset. In stage 1 of the framework,
the YOLOv3 convolutional neural network is used to accomplish
level one semantic hierarchy label, to count the number
of fishes and to measure the length of the detected fish. In
stage 2, the features from the images are extracted using the
VGG16 convolutional neural network. In stage 3, the stacked
generalization technique is implemented to reduce the generalization
error and to accomplish level two semantic hierarchy
label. The classification accuracy of the stack model is 94%.
The root mean square error of the fish length measurement is
1.23 cm. The accuracy in counting the number of fish depends
on the detection accuracy of the stage 1 model and the classification
accuracy of the stack models. Further, the results
can be improved by increasing the size and diversity of the
dataset.