Facial Emotion Recognition (FER) models increasingly govern automated clinical screening and human-computer interactions, yet their operational stability remains tethered to high-resolution data baselines. This paper presents an algorithmic audit of deep-convolutional emotion classification architectures (specifically leveraging the VGG-Face deployment framework via DeepFace) to evaluate performance stability under pixelation and compression constraints common to low-bandwidth developing regions. Utilizing a controlled pool of 400 neutral facial expressions from the FER-2013 repository, we contrast an uncorrupted control cohort against an experimental group subjected to programmatic resolution downscaling (12x12 native resolution back-interpolated to a 48x48 matrix). Our findings demonstrate an execution accuracy collapse from an 83.5% baseline classification rate in the high-resolution cohort down to 32.5% in the degraded experimental cohort. Crucially, error distribution was not random; instead, resolution degradation introduced a systematic negative affect bias, disproportionately misclassifying neutral resting expressions as Sad (23.5%), Fear (20%), and Angry (9%) . These empirical results reveal severe deployment inequities for emotion-driven healthcare applications in low-bandwidth global infrastructures and propose optimization protocols for adaptive multi-scale training datasets.
Naiyya Thapa (Sun,) studied this question.