针对传统去雾算法需要人工提取特征,对比度低、信噪比低等问题,提出一种基于多特征融合的卷积神经网络去雾算法。利用卷积神经网络算法模拟人类视觉系统对雾天图像进行层次化处理,实现自动提取特征。算法采用直接从雾天图像到清晰无雾图像映射的学习方式,该映射由特征提取、多尺度特征融合和浅层深层特征融合联合实现。多尺度特征融合提升网络对图像细节的重建,浅层深层特征融合则将浅层卷积得到的轮廓信息和深层卷积得到的细节信息进行融合,提升去雾重建的整体效果。实验结果表明,相比于单一尺度网络,多特征融合网络的峰值信噪比提高了1.280 dB。本文算法对自然雾天图像去雾效果明显,细节信息和对比度均优于其他算法,为去雾方法的研究提供了新思路。

We propose a convolutional neural network defogging algorithm based on multi-feature fusion to overcome the problem of manual feature extraction, low contrast, and low signal-to-noise ratio in traditional defogging algorithms. The convolution neural network simulates the human visual system to hierarchically process the fog images and automatically extract image features. The algorithm adopts a learning method of the direct mapping from the hazing image to the clear defogging image, which includes feature extraction, multi-scale feature fusion, and shallow and deep feature fusion. Multi-scale feature fusion helps to rebuild details of the image. Shallow and deep feature fusion combines the contour information obtained by shallow convolution with the detail information obtained by deep convolution to enhance the overall effect of defogging. The experimental results show that the peak signal to noise ratio of the multi-feature fusion network increases by 1.280 dB compared with the single-scale network. The proposed algorithm has obvious defogging effect on natural fog image and superior detail information and contrast compared with other algorithms, which provides a new idea for defogging methods.