Lightweight Design of Image Denoising based on TNet mini

Image denoising is vital in fields like autonomous driving and medical imaging to reduce noise while preserving quality. Traditional methods blur edges, and deep learning is computationally costly. This paper proposes TNet mini, a convolutional neural network based on denoising algorithm that balances denoising performance and efficiency. The algorithm introduces a network model that can be adapted to different application scenarios by selecting models of varying sizes. After training, the model is quantized from 32-bit to 8-bit, reducing storage requirements by 73% with only a slight decrease in denoising quality (PSNR drops by 0.215). The algorithm outperforms typical deep learning models in both performance and speed, making it suitable for deployment on edge devices. To address the high power consumption issues of commonly used convolutional neural network inference hardware, this paper designs a dedicated hardware accelerator for TNet mini network inference. The accelerator achieves 38.5135 GOPS/W energy efficiency, outperforming both the 2080Ti GPU (7.1429 GOPS/W) and i5-12400f CPU (2.5281 GOPS/W), offering up to 15.23 times better efficiency than the i5-12400f and 5.39 times better than the 2080Ti.