We propose a method to restore aspect ratio distortion of images using convolutional neural network (CNN). The “aspect ratio,” which is focused on this research, means degree of horizontal stretching of images. Indeed an image can be distorted by vertical or horizontal stretching, which does not maintain the aspect ratio. In the proposed method, we construct an aspect ratio estimator whose input is a (possibly distorted) image and output is a scalar value of aspect ratio. Since estimation of aspect ratio from image can be regarded as regression problem, we modeled the estimator by CNN. Once we have a reliable estimate of aspect ratio of an image, the restoration can be done straightforwardly by inverse stretching. In the experiments, we evaluated performance of the model trained on Pascal VOC natural image dataset. Our method can precisely restore the distortion within $1.4\%$ of stretch from original images on average, which outperforms average human performance (i.e., about $13\%$). In terms of accuracy, $99.86\%$ of distorted images are successfully restored. We also propose training methods to enhance the robustness of the CNN against particular types of disturbance.

中文翻译：

---

利用卷积神经网络恢复自然图像的宽高比畸变

我们提出一种还原方法 图像的宽高比失真使用卷积神经网络（CNN）。专注于这项研究的“长宽比”是指图像的水平拉伸程度。实际上，图像可能会因垂直或水平拉伸而失真，而无法保持纵横比。在提出的方法中，我们构造了一个长宽比估计器，其输入是（可能是失真的）图像，而输出是长宽比的标量值。由于从图像估算纵横比可以看作是回归问题，因此我们通过CNN对估算器进行建模。一旦我们有了图像长宽比的可靠估计，就可以通过逆拉伸直接完成恢复。在实验中，我们评估了在Pascal VOC自然图像数据集上训练的模型的性能。我们的方法可以精确地恢复内部的失真$ 1.4 \％$ 平均而言，其原始图像的拉伸度要比人类的平均水平高（例如，大约 $ 13 \％$）。在准确性方面，$ 99.86 \％$的失真图像已成功还原。我们还提出了训练方法，以增强CNN抵抗特定类型干扰的鲁棒性。