基于深度学习的手术器械视觉图像高斯与椒盐噪声去除方法研究

Deep learning-based Gaussian and pepper noise removal method for visual images of surgical instruments

目的:为消除手术器械视觉图像中的高斯与椒盐噪声,并恢复图像的细节特征,提出一种基于深度学习的手术器械视觉图像高斯与椒盐噪声去除方法。方法:首先,构建由多特征融合编码器解码器网络、注意力引导网络和细节恢复渐进式网络3个部分组成的轻量级多任务渐进式网络,使用多特征融合编码器解码器网络预测视觉图像中的噪声信息,并从原图像中去除,使用注意力引导网络进一步去除图像中的残留噪声,使用细节恢复渐进式网络对去噪图像的底层细节特征进行恢复。其次,对轻量级多任务渐进式网络进行轻量化设计,将细节恢复渐进式网络中的部分常规卷积替换为深度可分离卷积。最后,在公开的CBSD68、Kodak24数据集和自建的手术器械噪声数据集上进行去噪实验,比较基于轻量级多任务渐进式网络的去噪方法与经典去噪方法的去噪效果,ResNet-18模型和ResNet-34模型对采用轻量级多任务渐进式网络去噪后的图像的分类准确率,并分析轻量化设计前后的算力和内存占用情况。结果:在公开数据集上,所提出的方法较经典的去噪方法取得了更好的去噪效果。在手术器械噪声数据集上,ResNet-18模型和ResNet-34模型对采用轻量级多任务渐进式网络去噪后的图像分类准确率更高。经过轻量化设计的去噪方法的参数量和浮点运算数(floating point operations,FLOPs)分别减少了约27.27%和29.81%。结论:基于深度学习的手术器械视觉图像高斯与椒盐噪声去除方法具有优秀的手术器械视觉图像去噪性能,且具有更少的算力消耗和内存占用。

Objective To propose a deep learning-based method for removing Gaussian and pepper noises of the surgical instrument visual images so as to recover the detailed features of the images.Methods A lightweight multi-task progressive network was constructed involving in a multi-feature fusion encoder-decoder network,an attention-guided network and a detail-recovery progressive network,which used the multi-feature fusion encoder-decoder network to predict and eliminate the noise information in the visual images,the attention-guided network to remove the residual noise and the detail-recovery progressive network to restore the underlying detail features of the denoised images.Some of the regular convolutions in the detail recovery progressive network were replaced with depth separable convolutions to realize lightweight design of the network constructed.Denoising experiments were conducted on the publicly available CBSD68 and Kodak24 datasets and the self-constructed surgical instrument noise dataset so as to compare the denoising effects of the network constructed and the traditional methods and the classification accuracies of ResNet-18 model and ResNet-34 model for the denosied images by the network and to analyze computing power and memory usage before and after the lightweight design.Results The network constructed gained better denoising effect than the classical methods for publicly available datasets,and ResNet-18 model and ResNet-34 model had higher accuracies when used to classify the images denoised by the network for the self-constructed surgical instrument noise dataset.Lightweight design had the parameter number and floating point operations(FLOPs)decreased by approximately 27.27%and 29.81%,respectively.Conclusion The proposed lightweight multi-task progressive network behaves well in denoising surgical instrument visual images with reduced computating power consumption and memory usage.