基于注意力机制的双向生成对抗压缩网络

Bidirectional Generation Adversarial Compression Network Based on Attention Mechanism

针对现有的图像超分辨率算法网络模型参数量大、计算复杂度高、前向推理过程中耗时长等问题,将深度可分离卷积层引入双向对抗生成网络模型中,同时为了保证双向生成对抗网络的精度,在下采样网络中引入混合注意力机制,以保证模拟生成的低分辨率图片更加贴近现实。在i78700 CPU上对Urban100测试集的图像放大4倍,所提算法的重建速度对比SRGAN算法提升了近5倍。在通用的测试数据集DIV2K中,将所提方法分别与Bicubic、SRCNN、SRGAN、ESRGAN等经典方法做实验对比,实验结果表明所提出的算法与其他算法相比在PSNR客观评价指标上平均提升了约0.7 dB,SSIM指标提升了约2.36%。

In view of the problems of the large network model parameters,high computational complexity,and long time-consuming process of forward inference in the existing image super-resolution algorithm,this paper introduces the deep separable convolutional layer into the two-way confrontation generation network model,at the same time,in order to ensure the accuracy of the two-way generation against the network,a mixed attention mechanism is introduced in the down-sampling network to ensure that the lowresolution images generated by the simulation are closer to reality.On the i7-8700 CPU,the image in Urban100 test set is enlarged 4 times.The reconstruction speed of the algorithm in this paper is improved by nearly 5 times compared with the SRGAN algorithm.In the general test data set DIV2 K,the method of this paper is compared with Bicubic,SRCNN,SRGAN,ESRGAN and other classic methods.The experimental results show that the algorithm proposed in this paper has improved the average PSNR objective evaluation index compared with other algorithms about 0.7 dB,SSIM indicator increased by about 2.36%.