抽样切分卷积实现跨尺度特征融合及内镜图像去模糊

Sampling slice convolution for cross-scale feature fusion and endoscopic image deblurring

针对内镜图像去模糊过程中语义信息难以提取和细节纹理重建困难的问题,设计了一种新的抽样切分卷积,并将其应用于跨尺度特征融合过程中:通过等间隔抽样将大尺度特征无损切分成小尺度特征块,再与小尺度特征进行卷积融合。过程中大尺度特征的所有值都参与了特征融合,避免了细节信息的丢失;未对小尺度特征进行插值,避免了语义信息的模糊。为进一步实现特征互补,设计了特征交互融合模块,先用语义特征激活细节特征,再将两者融合。针对内镜图像亮通道、中间通道和暗通道的特征差异性设计了梯度重建和频域重建损失函数,提升了重建图像的锐度。在EAD和Kvasir-SEG数据集上,该算法的PSNR分别达到32.88 dB和33.01 dB,SSIM分别达到0.972和0.973。实验结果表明,该算法的性能优于主流去模糊算法,视觉上重建图像的纹理更清晰,且未产生伪影。

To solve the challenges of semantic information extraction and texture reconstruction in the process of endoscopic image deblurring,this paper designed a new sampling slice convolution(SSC)and applied it to the cross-scale feature fusion process.It divided the large-scale features into small-scale feature blocks losslessly by sampling at equal intervals,and then fused these feature blocks with small-scale features through convolution.All values of large-scale features participate in the feature fusion process,which could avoid the loss of detailed information.There was no interpolation operation on small-scale features,which could avoid the blurring of their semantic information.This paper proposed a feature interaction fusion(FIF)mo-dule,which used semantic features to activate detailed features,and then fused the two to achieve feature complementarity.This paper designed gradient reconstruction and frequency domain reconstruction loss functions for the feature differences of the bright channel,middle channel,and dark channel of endoscopic images to improve the sharpness of reconstructed images.Experiments on EAD and Kvasir-SEG datasets show that the PSNR of the algorithm reaches 32.88 dB and 33.01 dB,respectively,and the SSIM reaches 0.972 and 0.973,respectively.The experimental results show that the performance of the proposed algorithm is better than that of the mainstream deblurring algorithms,and the texture of the reconstructed image is visually clearer and does not produce artifacts.