基于改进型MBLLEN网络的内窥镜图像增强算法研究

Research on Endoscopic Image Enhancement Algorithm Based on Improved MBLLEN Network

针对内窥镜场景,亮度与细节增强算法可能会引入色调改变和图像噪声问题,提出改进型MBLLEN(Multi-Branch Low-Light Enhancement Network)网络对内窥镜图像进行增强。首先,采用改进型MBLLEN网络对内窥镜图像亮度进行增强,减少特征提取卷积层,并采用跳连接方式,将增强模块替换成U-Net结构;其次,引入Hessian矩阵特征值,对图像的全局线性结构做细节后处理增强。消融实验结果表明,改进结构在内窥镜场景下相比于原结构更具适应性,并且验证集的增强结果的平均PSNR值为26.829,SSIM值为0.868,DVBV值为131.372,该算法的指标结果与6种主流算法比较,均排名前列。由此可见,本研究算法在保证内窥镜图像质量的基础上,有效地提高了亮度与细节表现。

In response to the issues of hue alteration and image noise introduced by brightness and detail enhancement algorithms in endoscopic scenarios,this study proposes an improved MBLLEN(Multi-Branch Low-Light Enhancement Network)for enhancing endoscopic images.Firstly,the improved MBLLEN network is used to enhance the brightness of endoscopic images.The enhanced network reduces the number of convolutional layers for feature extraction and employs skip connections for linking.The enhancement module is replaced with a U-Net structure.Secondly,Hessian matrix eigenvalues are introduced for post-processing to enhance the global linear structure of the image details.The results of ablation experiments show that the improved structure has enhanced adaptability in endoscopic scenarios compared to the original structure.The average PSNR(Peak Signal-to-Noise Ratio)value of the enhancement results on the validation set is 26.829,the SSIM is 0.868,and the DVBV value is 131.372.The performance metrics are among the top when compared with seven mainstream algorithms.This demonstrates that the proposed algorithm effectively improves brightness and detail representation while ensuring the quality of endoscopic images.