College of Computer and Information, Hohai University, Nanjing 211100, China.
School of Engineering, University of Guelph, Guelph, ON N1G 2W1, Canada.
Sensors (Basel). 2018 Oct 22;18(10):3587. doi: 10.3390/s18103587.
In a traditional convolutional neural network structure, pooling layers generally use an average pooling method: a non-overlapping pooling. However, this condition results in similarities in the extracted image features, especially for the hyperspectral images of a continuous spectrum, which makes it more difficult to extract image features with differences, and image detail features are easily lost. This result seriously affects the accuracy of image classification. Thus, a new overlapping pooling method is proposed, where maximum pooling is used in an improved convolutional neural network to avoid the fuzziness of average pooling. The step size used is smaller than the size of the pooling kernel to achieve overlapping and coverage between the outputs of the pooling layer. The dataset selected for this experiment was the Indian Pines dataset, collected by the airborne visible/infrared imaging spectrometer (AVIRIS) sensor. Experimental results show that using the improved convolutional neural network for remote sensing image classification can effectively improve the details of the image and obtain a high classification accuracy.
在传统的卷积神经网络结构中，池化层通常采用平均池化方法：非重叠池化。但是，这种条件导致提取的图像特征之间存在相似性，特别是对于连续光谱的高光谱图像，这使得提取具有差异的图像特征更加困难，并且图像细节特征容易丢失。这一结果严重影响了图像分类的准确性。因此，提出了一种新的重叠池化方法，即在改进的卷积神经网络中使用最大池化来避免平均池化的模糊性。所使用的步长小于池化核的大小，以实现池化层输出之间的重叠和覆盖。本实验选择的数据集是由机载可见/红外成像光谱仪（AVIRIS）传感器采集的印度松树数据集。实验结果表明，使用改进的卷积神经网络进行遥感图像分类可以有效地提高图像的细节，并获得较高的分类精度。
