An efficient real-time pipeline for flare removal on mobile devices

In recent years, smartphones have incorporated increasingly advanced camera systems. However, their compact form factors restrict the use of complex lens designs, making lens flare artifacts particularly problematic in low-light photography. We propose an efficient two-stage software pipeline for real-time flare removal on mobile devices. The system first employs a binary classifier based on EfficientNet-V2-S, achieving 91% accuracy for flare detection, followed by a Multi-Frequency Deflare Network (MFD-Net) achieving 25.9 dB PSNR and 0.93 SSIM for flare removal. The detection module executes in 67–140 ms on mobile devices, while the full pipeline processes 1024×1024 images within 0.5–2 seconds. Our models are trained on synthetic and real-world datasets and deployed using ONNX Runtime for cross-platform execution with NPU and GPU acceleration, achieving inference times as low as 16 ms on desktop GPUs.