基于重影卷积的轻量化遥感图像超分辨重建

Lightweight super-resolution reconstruction of remote sensing images based on Ghost convolution

为了解决基于深度学习的遥感图像超分辨率重建算法模型复杂度高、纹理细节重建不准确的问题,本文提出了基于重影卷积的信息多蒸馏网络。采用重影卷积(Ghost convolution)代替传统卷积消除冗余,降低模型复杂度;通过引入改进的高频注意力块(Improved high-frequency attention block,IHFAB)提高网络对图像中高频成分的特征捕获能力,优化网络对纹理轮廓等细节的重建能力。实验结果表明,本文提出的方法相较于残差特征蒸馏网络(Residual feature distillation network,RFDN)等参数量明显降低,相较于蓝图可分离残差网络(Blueprint separable residual network,BSRN),在2倍、3倍和4倍放大下峰值信噪比分别提升0.33、0.30和0.11 dB,结构相似度分别提升0.017、0.005和0.007。

In order to solve the problems of high model complexity and inaccurate reconstruction of texture details in the deep learning-based super-resolution reconstruction algorithm for remote sensing images,an information multi-distillation network is proposed based on Ghost convolution.Ghost convolution is used instead of traditional convolution to eliminate redundancy and reduce the model complexity.The feature capture ability of the network for high-frequency components in the image is improved through the introduction of Improved high-frequency attention block(IHFAB)to optimize the reconstruction ability of the network for details such as texture contours.The experimental results show that the method proposed in this paper significantly reduces the isoparametric number compared to Residual feature distillation network(RFDN),improves the peak signal-to-noise ratio by 0.33,0.30 and 0.11 dB at 2×,3×and 4×scaling,and improves the structural similarity by 0.017,0.005 and 0.007,respectively,compared to the Blueprint separable residual network(BSRN).