基于BP-Net网络结构的图像去雾算法研究

Image Dehazing Algorithm Based on BP-Net Network Structure

神经网络已被大量应用于图像去雾领域,并取得了较好的效果。但是一般用于去雾的神经网络的深度较深且结构复杂,不利于其在嵌入式平台上的部署。提出了一种基于基础块堆叠的架构,BP-Net(Block Piled Network)。并分别尝试用残差块与信息蒸馏网络(IDN)的信息蒸馏块作为网络中的基础块。此外,采用感知损失函数进行网络训练。训练结果在SOTS测试集上的平均PSNR与SSIM值分别达到了33.15 dB与0.9772。此外,在大幅缩减基础块数量后,网络去雾效果也有良好表现。

In recent years,neural networks have been widely used in the image dehazing and achieved good results.However,the neural network applied in image dehazing always has a deeper depth and a more complex structure,which is a disadvantage while appling in a embedded system platform.A structure called BP-Net(Block Piled Network),which is based on the piling basic block structure is proposed.The residual network(ResNet)and the information distillation module of information distillation network are used as the basic block respectively.The perceptual loss is also adopted as the training loss of our network.Testing on the SOTS dataset,the averange PSNR has achieved 33.15 d B while SSIM has achieved0.9772.Besides,the number of the basic blocks are also reduced and a good result is achieved.