基于多输入对抗网络的水下图像增强
首发时间:2019-10-22
摘要:由于水体吸收和散射的影响,水下图像往往存在纹理特征模糊、对比度降低以及颜色失真等现象。为提高水下图像的视觉质量,本文提出一种基于多输入对抗网络水下图像增强。首先,利用经典的DCP算法对原始图像进行去雾,针对上去雾图像分别使用白平衡(WB)和对比度受限的自适应直方图均衡化(CLAHE)算法,得到颜色矫正和对比度增强图像;然后,利用编码器-解码器学习原始图像、颜色校正和对比度增强图像之间的特征差异的置信度图,为减少WB和CLAHE算法引入的伪影和细节模糊,引入两个子网络分别滤除干扰信息,保留重要特征信息;最后,通过对每个子网络输出的重要特征进行门融合操作,利用判别网络和参数固定的VGG-19网络进行训练以达到最优解,获得增强的水下图像。对增强后的水下图像进行测试与评估,实验结果表明,增强后的水下图像色彩鲜明和细节清晰,该网络对视觉效果有显著提升。
For information in English, please click here
Multi-input adversarial network for underwater image enhancement
Abstract:Due to the influence of water absorption and scattering, underwater images often have blurred texture features, reduced contrast, and color distortion. In order to improve the visual quality of underwater images, the paper proposes an underwater image enhancement based on multi-input adversarial network. First, the original image is defogged using the classical DCP algorithm, and white balance (WB) and contrast-limited adaptive histogram equalization (CLAHE) algorithm are used for the defogged image to obtain color correction and contrast enhancement images; Then,encoder-decoder is used to learn the confidence map of the difference between the original image, the color correction and the contrast-enhanced image. To reduce the artifacts and detail blur introduced by the WB and CLAHE algorithms, two sub-networks are introduced to filter the interference informationseparately,The important feature information is retained; Finally, the gate fusion operation is performed on the important features of each sub-network output,discriminant network and VGG-19 network with the fixed parameter is used for training to achieve the optimal solution, the enhanced underwater image is obtained. The enhanced underwater image was tested and evaluated. The experimental results show that the enhanced underwater image has clear colors and clear details, and the network has a significant improvement in visual effects.
Keywords: Generate adversarial network Fusion Dense block Underwater image enhancement
引用
No.****
动态公开评议
共计0人参与
勘误表
基于多输入对抗网络的水下图像增强
评论
全部评论0/1000