基于视觉Transformer的多损失融合水下图像增强网络

Underwater image enhancement network based on visual Transformer with multiple loss functions fusion

由于水中存在光的吸收和散射现象,水下机器人拍摄到的图像存在颜色失真和对比度降低的问题。针对水下图像存在的质量退化现象,提出了一种基于视觉Transformer的多损失融合的方式训练水下图像增强网络。图像增强网络采用编码与解码的结构,可以采用端到端的方式进行训练。将多损失的线性组合作为总体优化目标,有效地更新水下图像增强网络的参数,包括像素损失、结构损失、边缘损失和特征损失。在两个大型水下数据集上进行了量化实验,并与7种水下图像增强算法进行对比。以峰值信噪比和结构相似性为有参考评估指标,以水下评估指标为无参考评估指标进行实验。实验结果表明,提出的水下图像增强网络能够有效地解决图像的颜色失真与对比度降低问题。

Due to the absorption and scattering of light in water,the images captured by underwater robots suffer from color distortion and reduced contrast.Aiming at alleviating the quality degradation phenomenon of underwater images,an underwater image enhancement network based on vision Transformer that be trained with multiple losses fusion strategy was proposed.The image enhancement network adopted an encoder-decoder architecture,and could be trained in an end-to-end manner.In order to effectively update the parameters of the network for enhancing underwater images,a linear combination of various losses was adopted as the overall optimization objective,including pixel loss,structure loss,edge loss and feature loss.Quantitative experiments were carried out on two large underwater datasets,and the proposed underwater image enhancement network was compared with 7 underwater image enhancement algorithms.The full reference evaluation metrics peak signal-to-noise ratio and structural similarity were calculated in experiment,and the non-referenced metric underwater image quality measure was also computed.The experimental results showed that the proposed underwater image enhancement network could effectively deal with color distortion and contrast reduction.