基于上下文编码器的图像修复算法

Image inpainting algorithm based on context-encoder

基于数学的传统模型在纹理学习上的效果比较好,但是在图像内容和语义上有所缺失。近年来,随着深度学习的兴起,将深度学习应用于图像修复的方法十分热门,对各种残缺图像的修补也取得了较好的结果。本文使用的基于上下文编码器的图像修复方法,结合自编码器(AE)和生成对抗网络(GAN)。其中,使用AE进行图像特征的学习,生成待修补区域的预测图;使用GAN的对抗学习来优化模型;使用由重建损失和对抗损失组成的联合损失函数。经过训练后的卷积神经网络(CNN),能够根据图像周边的像素特征,对丢失的区域进行合理推断,生成缺失部分。最后,本文对比了使用联合损失函数和单独使用重建损失与对抗损失函数的效果,并将本文算法与另外2种模型对比,采用客观评价指标,即平均绝对误差(MAE)、峰值信噪比(PSNR)以及结构相似性(SSIM)进行评价,其结果表明本模型在修复残缺图像时效果均较好,普适性更强。

The traditional models based on mathematics have good results in texture learning,but they lack in image content and semantics.In recent years,with the rise of deep learning,the application of deep learning to image repair is very popular,and the repair of various mutilated images has also obtained good results.This paper uses the context-encoder-based image repair method,combining auto encoder(AE)and generative adversarial networks(GAN).AE is used to learn the image features,generate the prediction map of the area to be repaired,and confrontation learning of GAN is used to optimize the model,and a joint loss function consisting of reconstruction adversarial loss is employed.The trained convolutional neural network(CNN)can reasonably infer the missing region according to the pixel characteristics around the image and generate the missing part.Finally,the mean absolute error(MAE),peak signal-to-noise ratio(PSNR)and structural similarity index(SSIM)are used for evaluation to compare the effectiveness of using joint loss functions with using reconstruction loss and adversarial loss separately and to compare the algorithm with the other two models.The results show that the proposed model has better effect and better universality in repairing mutilated images.