ANDRIANTO, ALVIEN (2026) IMPLEMENTATION OF LIGHTWEIGHT ATTENTION BASED CNN WITH CLAHE PRE-PROCESSING FOR ROBUST FACIAL EXPRESSION RECOGNITION UNDER LOW-LIGHT AND PARTIAL OCCLUSION CONDITIONS. S1 thesis, UNIVERSITAS KATOLIK SOEGIJAPRANATA SEMARANG.

Preview	Text 22.K1.0035 - ALVIEN ANDRIANTO BUDIHARJO-COVER_a.pdf Download (858kB) | Preview
	Text 22.K1.0035 - ALVIEN ANDRIANTO BUDIHARJO-ISI_a.pdf Restricted to Registered users only Download (1MB)
Preview	Text 22.K1.0035 - ALVIEN ANDRIANTO BUDIHARJO-DAPUS_a.pdf Download (710kB) | Preview
	Text 22.K1.0035 - ALVIEN ANDRIANTO BUDIHARJO-LAMP_a.pdf Restricted to Registered users only Download (704kB)

Abstract

Facial Expression Recognition (FER) systems often fail in real world environments due to poor lighting and covered facial features. To solve this, we propose a highly efficient pipeline that combines Contrast Limited Adaptive Histogram Equalization (CLAHE) to recover details in dark images, and a Convolutional Block Attention Module (CBAM) inside a custom lightweight CNN to maintain focus when parts of the face are hidden. We evaluated this model using the CK+ dataset under controlled stress tests. During severe partial occlusion (like wearing a mask), the proposed model successfully maintained an 83.76% accuracy, suffering only a 10.15% drop rate, whereas a standard CNN crashed to 48.22%. In extreme low-light conditions (30% illumination), CLAHE pre-processing salvaged the model’s recognition ability, achieving 40.10% accuracy compared to the baseline's 17.77%. Furthermore, Grad-CAM visualizations confirmed that our attention mechanism actively shifts computational focus from blocked areas to visible facial muscles. While this aggressive training caused a minor 4.5% drop in peak accuracy under perfect conditions (from 98.48% to 93.91%), the resulting structural resilience is a necessary and highly valuable trade-off. Finally, the entire architecture is extremely efficient, requiring only 686,049 parameters (8.1 MB) with an inference speed of 75.99 ms per image. This proves that integrating adaptive pre-processing and spatial attention creates a robust FER system that is fully optimized for real-time edge devices.

Actions (login required)

View Item