CEPDNet: a fast CNN-based image denoising network using edge computing platform

CEPDNet: a fast CNN-based image denoising network using edge computing platform

연구 분야: Networking

학회: The Journal of Supercomputing

Edge-preserving denoising is important in image analysis such as image segmentation, object detection and biometric recognition. However, most of the existing image denoising methods suffer from the smoothness of high-frequency edge details and high computational complexity in the inference stage. Aiming at this issue, we propose a CNN and tensorRT-based edge-preserving denoising network and complete the deployment and inference acceleration on the edge computing platform through model optimization. Specifically, first, a parallel edge extraction module including wavelet transform and threshold unit is designed to generate the de-redundant edge information from the noisy image. Then, the edge is used as a prior to generate affine transform parameters to improve the weight distribution between the noise features and the edge features, allowing the network to focus more on learning the noise features other than the high-frequency information at the edges. In the deep feature extraction stage, the spatial neighborhood information of feature points is obtained via a newly constructed multi-scale attention module, which is employed to activate the deep features and further improve the ability of the network to extract noise features. Finally, the computational efficiency of the model in the inference stage is further improved through the optimization of the training model with lightweight network design and TensorRT. Extensive experimental results show that the proposed method can effectively preserve image edge details while maintaining a higher inference speed, i.e., a denoising frame rate of about 40fps at the image resolution of 256 × 256.