A Lightweight Dual-Branch CNN with Frequency Domain Perception Loss for Image Denoising

Lightweight real-time image denoising is crucial for resource-constrained edge devices, yet existing compact convolutional neural networks (CNNs) often lose high-frequency details due to limited capacity and the absence of explicit frequency-domain supervision. This paper proposes a 0.18M-parameter dual-branch denoising network driven by a novel Frequency Domain Perception Loss (FDPL). The architecture decouples noise removal and detail recovery via a low-frequency branch composed of four Residual-in-Residual Dense Blocks (RRDB) and a high-frequency branch with two Residual Channel Attention Blocks (RCAB). The composite loss combines brightness-aware Mean Square Error (MSE), Visual Geometry Group 19-layer (VGG19) perceptual loss, and the FDPL, which applies a circular mask in the frequency domain (high/low weights of 0.8/0.2) with linear warm-up. Extensive experiments on Berkeley Segmentation Dataset with 68 images (BSD68), its color counterpart (CBSD68), and Urban100 (100 high-resolution urban scene images) demonstrate the model’s effectiveness: the FDPL consistently improves Peak Signal-to-Noise Ratio (PSNR) by 0.37–0.69 dB across all datasets, pushing the peak performance to 31.37 dB on BSD68, which surpasses classic Denoising Convolutional Neural Network (DnCNN) by 5.28 dB. When further evaluated against recent state-of-the-art lightweight models on the Berkeley Segmentation Dataset (500 images) (BSD500) benchmark under both low and moderate noise levels, our method achieves highly competitive results–31.15 dB PSNR at σ = 12.75, approaching the much larger Real Image Denoising Network (RIDNet) (32.34 dB) while being over five times smaller and exhibiting 7.4× faster inference. It surpasses SwinIR-light by 0.28 dB with 33× lower latency at σ = 12.75 and maintains comparable accuracy with over 30× speedup at σ = 25. With only 0.18M parameters and 3.29 ms Central Processing Unit (CPU) inference per 64 × 64 patch, the model is readily deployable on edge devices. Ablation studies confirm the importance of both the dual-branch design and the FDPL. These results highlight that the proposed architecture and loss function enable an ultra-lightweight CNN to rival significantly heavier models, making it particularly suitable for real-time denoising in resource-limited applications such as mobile imaging, surveillance, and on-device preprocessing.