基于记忆增强Transformer的单幅散焦图像去模糊OA
Single-image defocus deblurring based on a memory-augmented Transformer
现有基于 Transformer 的图像复原方法虽然展示了去除散焦模糊的潜力,但由于自注意力机制的全局特性,模型在处理过程中会丢失图像细节特征,限制了模糊去除的效果.针对这一问题,提出一种基于记忆增强 Transformer 的单幅散焦图像去模糊方法.该方法在特征提取与重建的各阶段引入记忆增强Transformer 块,通过动态融合记忆特征与当前查询特征,以增强模型对细节信息的保持能力.具体而言,记忆特征由图像浅层特征初始化,并在网络各阶段自适应调整,以融合全局上下文与局部细节.在 Trans-former 块内部,记忆机制通过动态分配查询特征与记忆特征的权重,生成信息更丰富的融合特征,同时更新记忆状态.在 DPDD 和 LFDOF 等多个数据集上的实验结果表明,所提方法在散焦模糊去除方面表现优异.在 DPDD 数据集上 PSNR 达到 26.23 dB,SSIM 为 0.813;在 LFDOF 数据集上 PSNR 达到 31.59 dB,SSIM 为 0.908,LPIPS 为 0.119,各项指标均优于现有主流方法,且展现出良好的泛化能力.
Existing Transformer-based image restoration methods have demonstrated potential in removing defocus blur,but the global self-attention mechanism often leads to the loss of image details,limiting the effectiveness of defocus deblurring.To address this issue,this paper proposes a single image defocus deblurring method based on a memory-augmented Transformer.The method introduces memory-augmented Transformer blocks throughout the feature extraction and reconstruc-tion stages,dynamically fusing memory features with current query features to enhance the model's ability to preserve image details.Specifically,the memory features are initialized with shallow im-age features and adaptively adjusted at each stage of the network to integrate global context and local details.Inside the Transformer block,the memory mechanism dynamically allocates weights be-tween query features and memory features to generate more informative fused features while updating the memory representation.Experimental results on multiple datasets,including DPDD and LFD-OF,demonstrate that the proposed method achieves excellent performance in defocus deblurring.On the DPDD dataset,it reaches a PSNR of 26.23 dB and an SSIM of 0.813;on the LFDOF data-set,it achieves a PSNR of 31.59 dB,an SSIM of 0.908,and an LPIPS of 0.119.The proposed method outperforms existing mainstream methods on all metrics and exhibits good generalization abil-ity.
杨劲怀;杨小娟;陈静;何玉琦;林亦辰;郑千玲;牛玉贞
国网福建省电力有限公司信息通信分公司,福建 福州 350000国网福建省电力有限公司信息通信分公司,福建 福州 350000国网福建省电力有限公司信息通信分公司,福建 福州 350000福州大学计算机与大数据学院,福建 福州 350116国网福建省电力有限公司信息通信分公司,福建 福州 350000国网福建省电力有限公司信息通信分公司,福建 福州 350000福州大学计算机与大数据学院,福建 福州 350116
信息技术与安全科学
图像复原图像去模糊散焦模糊Transformer图像细节
image restorationimage deblurringdefocus deblurringTransformerimage de-tails
《福建师范大学学报(自然科学版)》 2026 (5)
13-24,12
国网福建省电力有限公司科技项目(52130M25000C)
评论