Hunger behaviour classification of Lates Calcarifer using machine learning for automatic demand feeder through image processing
The understanding and identification of fish hunger behaviour are non-trivial in the aquaculture industry. This thesis aims at classifying the hunger state of Lates Calcarifer via the integration of computer vision and machine learning. Prior to the classification of the hunger states, the hunger st...
Saved in:
Main Author: | |
---|---|
Format: | Thesis |
Language: | English |
Published: |
2019
|
Subjects: | |
Online Access: | http://umpir.ump.edu.my/id/eprint/31083/1/Hunger%20behaviour%20classification%20of%20lates%20calcarifer%20using%20machine%20learning%20for%20automatic%20demand.pdf |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Summary: | The understanding and identification of fish hunger behaviour are non-trivial in the aquaculture industry. This thesis aims at classifying the hunger state of Lates Calcarifer via the integration of computer vision and machine learning. Prior to the classification of the hunger states, the hunger state of the fish is identified through the k-means clustering technique and it was established that the hunger state could be demarcated into either ‘Hungry’ or ‘Satiated’. Upon the identification of the hunger state, significant features that could contribute towards the accurate classification of the states are investigated. The aforesaid features are analysed by the box plot analysis and the Principal Component Analysis (PCA). The established features are COG x, COG y and the moving summation of the pixel. Different machine learning models were investigated by incorporating the identified features, i.e., Discriminant Analysis (DA), Support Vector Machine (SVM) and k-Nearest Neighbours (k-NN) and it was demonstrated that the SVM trained model is able to classify up to 99.00%, suggesting that the developed system is viable for fish farming. A supplementary analysis was further carried out to understand the circadian rhythm of the fish by evaluating the time-series features. Different window sizes ranging from 0.5 min, 1.0 min, 1.5 min and 2.0 min coupled with the mean, maximum, minimum and variance for each of the distinctive temporal window sizes are investigated. PCA and PCA varimax rotation was employed in order to identify the best features through classifying it via SVM and k-NN. It was shown that the mean and variance of all temporal sizes are significant. In addition, the efficacy of different models based on the identified secondary features, namely DA, SVM, k-NN, Decision Tree (Tree), Logistic Regression (LR), Random Forest Tree (RF) and Neural Network (NN) are evaluated. It was found that the k-NN yielded the highest classification accuracy with 96.47% from the test sets. In order to further refine the k-NN model developed, hyperparameter optimization by means of Bayesian Optimization was carried out. Through the optimization process, the best hyperparameters that could attain a classification accuracy of 97.16% are the Standardized Euclidean distance metric with a k value of one. |
---|