Обмеження навчання моделей комп'ютерного зору для систем розпізнавання психоемоційних станів в авіаційній інфраструктурі

Description

Abstract (English)

Abstract. In the era of rapid digitalization within the aviation sector, the integration of artificial intelligence and computer vision systems is increasingly critical for infrastructure security and passenger management. However, the automated recognition of complex psycho–emotional states remains a multi–dimensional challenge, as contemporary neural network models frequently fail to interpret non–standard, context–dependent, or subtle affective expressions. This study critically evaluates the fundamental epistemological and algorithmic limitations of object detection architectures, specifically focusing on YOLOv8, within the high–stakes environment of aviation infrastructure.

Adopting a transdisciplinary STEAM–based (Science, Technology, Engineering, Art, Mathematics) educational framework, the research bridges the gap between biochemical emotional determinants, mathematical modeling, and humanistic inquiry. We implemented a specialized experimental project, utilizing a curated dataset of art history masterpieces as "critical stress tests" to evaluate the model’s capacity for emotional synthesis. Quantitative analysis revealed significant performance disparities across emotional classes, demonstrating high accuracy for overt, anthropomorphic patterns (e.g., "Happy" 𝑚𝐴𝑃@50 = 0,795) but critical failure in recognizing low– amplitude or concealed cognitive patterns (e.g., "Sad" 𝑚𝐴𝑃@50 = 0,260). Qualitative stress–testing – specifically utilizing Gustave Courbet’s Despair – exposed the phenomenon of "algorithmic reduction," where the system misclassifies existential crises as "surprise" (𝑐𝑜𝑛𝑓 = 0,63) due to its reliance on morphological markers rather than contextual synthesis.

We argue that the current reliance on discrete categorical labeling poses substantial risks to aviation security and passenger profiling. The study introduces the mathematical metaphor of the "inflection point" (𝑓′′(𝑥) = 0) to illustrate that AI must transition from static labeling to recognizing the non–linear dynamics of emotional trajectories. We conclude that successful implementation of emotion recognition in aviation environments necessitates a paradigm shift toward multimodal, context–aware, and Zero–Shot Learning architectures, moving beyond superficial pixel analysis toward an empathetic, cognitive synthesis of human behavior.

Beyond technical validation, this study underscores the pedagogical importance of project– based research in higher aviation education. The implementation of this framework transformed aviation education seekers from passive technology users into skilled analysts capable of critically evaluating the functional limitations of neural network models. The findings demonstrate that resolving the challenges of automated emotion recognition requires not only the development of multimodal, context–aware, and Zero–Shot Learning architectures but also a fundamental shift in professional training. Preparing future aviation specialists necessitates integrated educational programs that foster deep cognitive synthesis, ethical understanding of technology, and an ability to navigate the non–linear dynamics of human behavior. This approach ensures the responsible and reliable implementation of innovative profiling systems within the complex ecosystem of aviation operations.

Keywords: computer vision, YOLOv8, psycho–emotional state recognition, aviation infrastructure, higher aviation education, algorithmic reduction, STEAM education, artificial intelligence, security profiling, fine–tuning.