YOLOv8-FLEO: Fuzzy-Label Emotion Orthogonalization for Real-Time Facial-Expression Recognition Implementation

Facial Expression Recognition (FER) models frequently struggle with intrinsic emotional label ambiguities caused by overlapping facial action units, leading to high misclassification rates among confusable categories such as fear and disgust. Simultaneously, deploying advanced deep learning architectures enhanced with custom mathematical constraints onto resource-constrained embedded FPGA accelerators introduces severe operator fragmentation, where non-native operations trigger costly CPU-FPGA context switches. To address these challenges, this paper introduces YOLOv8-FLEO framework, a novel framework that integrates mutually orthogonal per-emotion subspaces to resolve feature manifold entanglement, combined with an innovative structural fold-out and post-fold fine-tuning deployment pipeline. The proposed approach eliminates DPU-hostile Gram–Schmidt operators from the network backbone, enabling seamless compilation into optimized Xilinx DPU subgraphs with zero non-native body operations. Comprehensive evaluations on the RAF-DB and FER2013 benchmarks demonstrate that YOLOv8-FLEO framework significantly elevates minority-class recall while maintaining high overall recognition accuracy (0.858 on RAF-DB). Furthermore, hardware implementation on the ZCU104 evaluation board targeting the DPUCZDX8G B4096 configuration reveals that the design operates securely on the compute-bound roofline plateau (∼266op/byte). Supported by INT8 Post-Training Quantization with negligible degradation (Δq≈0.004), the hardware accelerator achieves a sustained on-board throughput of ∼135FPS, a low latency of 7.2ms per frame, and an energy efficiency of ∼68mJ per inference, establishing a powerful paradigm for real-time, edge-optimized affective computing.