遥感影像要素提取的可变结构卷积神经网络方法

针对利用经典卷积神经网络提取遥感影像地物要素的方法中,模型容量受既定网络固有结构的限制而难以得到良好提取效果的问题,提出利用可变结构卷积神经网络的遥感影像要素提取方法。该方法将结构搜索与权重求解过程统一,在定义卷积神经网络架构的基础上将其中的关键结构作为变量,并以要素提取精度指标作为目标函数,利用遗传算法求解网络结构,最后以该网络为模型提取遥感影像中的目标要素。相关试验表明,可变结构卷积神经网络具备灵活的模型容量,对遥感影像中目标要素的提取效果良好。

Surface features extraction in remote sensing images based on architecture-variant CNN

To exceed limited capacity of established convolutional neural network(CNN) with fixed architecture in traditional surface feature extraction in remote sensing images, we propose a new feature extraction method based on architecture-variant convolutional neural network (AVCNN). In AVCNN, key units are variables and the performance of the unknown model become object function. That means architecture search is added before traditional weights solving. Genetic algorithm is introduced to search proper architecture and classical algorithm is used to solve unknown weights in the candidate CNN. The CNN with final architecture is used to extract the surface feature in remote sensing images. The experiment result shows that AVCNN has flexible capacity and performances well in surface features extraction in remote sensing images.