分层特征融合网络航拍图像超分辨率重建

Super-Resolution Reconstruction of Aerial Images Based on Hierarchical Feature Fusion Network

基于深度学习的图像超分辨率网络模型复杂度高,特征利用率较低,尤其是应用在复杂拍摄环境中的图像超分辨率重建,由于特征损失严重,最终重建的效果也较差。针对以上问题,提出分层特征融合图像超分辨率网络。引入对称式的分层结构,以增强不同层次图像特征的融合;使用更为密集的残差连接结构,减少局部残差损失,同时缓解梯度消失和梯度爆炸问题;在每个残差块中加入注意力机制,增强网络对图像高频信息的敏感度。为了验证算法在复杂环境中的效果,将模型应用于高空航拍图像超分辨率重建中。实验结果表明,所提算法相比于EDSR算法,在14个不同航拍图像环境中,尤其是复杂场景下的重建,平均PSNR提高了0.31 dB,效果显著。

Most of the current image super-resolution network models based on deep learning have high model complexity and low feature utilization,which leading to poor performance of final reconstruction due to severe feature loss,especially when they are applied in a complex shooting environment.In order to solve the problems mentioned above,a hierarchical feature fusion network(HFFN)for image super-resolution is proposed.Firstly,a symmetrical layered structure is intro-duced to enhance the fusion of image features at different levels.Secondly,it can not only reduce local residual loss but the gradient vanishing and gradient exploding can be effectively avoided by using a denser residual connected structure.Finally,by adding attention module to each residual block,the sensitivity of the model to high-frequency information of the image is enhanced.The proposed model has been used in the super-resolution reconstruction of high-altitude aerial photography environment for verifying its effectiveness.The experimental results show that the proposed algorithm has remarkable reconstruction performance and its average PSNR is increased by 0.31 dB in comparison with the EDSR algo-rithm when tested in 14 different aerial environments,especially the reconstruction of complex scenes.