Date of Award

1-1-2022

Thesis Type

masters

Document Type

Thesis (Restricted Access)

Divisions

fsktm

Department

Faculty of Computer Science & Information Technology

Institution

Universiti Malaya

Abstract

Real-world food datasets are not fixed, it is open-ended and dynamic, however, the novel machine learning methods for food recognition have poor performance in incremental learning datasets. If food samples and food categories continuous increase, these methods may need to train again from the beginning. This is time-consuming and occupies computational resources. My study proposed a multilabel classifier for this shortcoming, called Multi-Label Adaptive Reduced Class Incremental Kernel Extreme Learning Machine, the abbreviation is ARCIKELM-ML. We applied Inception-Resnet-V2 for food feature extraction and the Relief F method for feature ranking and selection. Then used ARCIKELM-ML for multi-label classification. In the framework, the hidden and output neurons corresponding to new labels are added and the classifier progressively remodels its structure like the new labels are introduced from the beginning of the training process. The experiment for food ingredients recognition is based on three standard benchmark datasets and evaluated on F1 score, Hamming Loss, Recall Score and Precision Score. Results showed that the proposed ARCIKELM-ML algorithm has good performance and meets the criteria of incremental learning

Note

Dissertation (M.A.) – Faculty of Computer Science & Information Technology, Universiti Malaya, 2022.

Share

COinS