Enhancing Hierarchical CNN Via Pathway Loss And Reference Forwarding Synergy

Adrus Mohamad, Tazuddin and Azizi, Abdullah and Zainal Rasyid, Mahayuddin (2023) Enhancing Hierarchical CNN Via Pathway Loss And Reference Forwarding Synergy. In: 2023 International Conference on Electrical Engineering and Informatics (ICEEI), 10-11 October 2023, Bandung, Indonesia.

PDF Enhancing Hierarchical.pdf Download (1MB)

Abstract

Computer vision applications have been widely used in critical areas however its reliability is hampered by low accuracy when dealing with small objects. As the internal information for small objects is limited by its pixels, addition of external information from the object context is important to determine object identity. Context features that are shared among objects can be used to group and label them hierarchically. It will then be used to improve classification by reducing the search space and linking the object’s general feature or function. Hierarchical-based CNN uses a hierarchy of labels with multi-level classifiers in classification of simpler coarser level before tackling more complex classification on finer level. However, the current implementation of hierarchical CNN keeps information between levels isolated from each other thus limiting the potential of its structure. Improvement is made towards the loss function with the addition of pathway loss where the hierarchy pathway from coarse to fine is considered on each classifier branch. Additionally, we introduce the reference forwarding function, which transfers crucial information learned from previous layers to the subsequent finer ones, thereby further improving the classification accuracy. Proof of concept experiments on CIFAR-100 dataset shows each of the proposed innovations improve accuracy compared to baseline model and the overall combinations synergize well with each other, achieving the best result with improvement of almost 8%.

Actions (For repository members only: login required)

View Item