基于改进RDN网络的无人机茶叶图像超分辨率重建

Super⁃resolution Reconstruction of Unmanned Aerial Vehicle Tea Images Based on Improved RDN Network

针对无人机搭建可见光传感器进行茶叶长势、病害等监测中因飞行高度影响图像分辨率的问题,本文提出了一种改进的残差密集网络(Residual dense network,RDN)用于无人机茶叶图像超分辨率重建。针对无人机茶叶图像纹理复杂的特点,以RDN为基线网络,在其结构中引入了残差组(Residual group,RG)模块,将多个残差通道注意力模块(Residual channel attention block,RCAB)组合在一起,通过引入注意力机制来区别对待不同的通道,关注无人机茶叶图像高频细节信息,从而提高网络的表征能力;同时设计了一个卷积长跳跃结构,利用带有卷积的远程跳跃连接,动态调整经过残差密集块(Residual dense block,RDB)后特征的权重,更好地利用无人机茶叶图像的分层特征信息,从而提升超分辨率重建图像的质量。实验结果表明,本文改进的RDN网络在无人机茶叶图像测试集上相较于其他算法表现更优,超分辨率重建后的图像具有更高的峰值信噪比和结构相似度,在4倍超分的情况下分别达到36.03 dB和0.9132,能够为茶叶智能化监测研究提供支持。

It is a relatively economical,flexible and time⁃effective method to build a visible light sensor for monitoring of tea growth and diseases,but the resolution of the image will be affected by the flying height of the UAV.Therefore,an improved residual dense network(RDN)for super⁃resolution reconstruction of UAV tea images was proposed.Specifically,in view of the complex texture of UAV tea images,taking RDN as the baseline network,residual group(RG)was introduced into its structure,combining multiple residual channel attention modules were combined together to treat different channels differently by introducing an attention mechanism,and paying attention to the high⁃frequency detail information of UAV tea images,thereby improving the representation ability of the network;at the same time,a convolutional long jump structure was designed,using the long⁃range skip connection with convolution,to dynamically adjust the weight of the feature after passing through the residual dense block(RDB),and making better use of the hierarchical feature information of the UAV tea image,thereby improving the super⁃resolution of the quality of reconstructed image.The experimental results showed that the improved RDN network performed better than other algorithms on the test set of UAV tea images,and the super⁃resolution reconstructed images had higher peak signal⁃to⁃noise ratio and structural similarity.In the case of quadruple super resolution,it can reach 36.03 dB and 0.9132,respectively,which can provide support for the follow⁃up research of tea intelligent monitoring.