摘要
针对红外图像信噪比低,对比度不足的问题,本文旨在红外图像增强,在对已有的比较流行的图像增强技术的研究的基础上,提出了引入无参数的注意力机制模块(SimAM)和使用加入循环残差结构(RRB)的U-Net代替的原始RetinexNet增强网络的图像增强算法。首先,通过在分解网络部分引入注意力机制提高空间特征提取能力,得到图像的光照分量和反射分量。其次,将光照分量和反射分量送入加入循环残差结构的U-Net增强网络,得到增强后的光照分量。最后,将增强后的光照分量和去噪后的反射分量相乘输出增强的红外图像。实验结果表明,与传统Retinex算法和原始RetinexNet算法相比,该算法能够有效提高对比度,丰富图像细节纹理,提高了红外图像的质量。
Aiming at the problems of low signal-to-noise ratio and insufficient contrast of infrared image, this paper aims at infrared image enhancement. On the basis of the research on the existing popular image enhancement technologies, an image enhancement algorithm is proposed, which introduces a parameterless attention mechanism module (SimAM) and uses U-Net with a cyclic residual structure (RRB) to replace the original RetinexNet enhancement network. Firstly, the attention mechanism is introduced into the decomposition network to improve the spatial feature extraction ability, and the illumination component and reflection component of the image are obtained. Secondly, the light component and reflection component are fed into the U-Net enhancement network with cyclic residual structure to obtain the enhanced light component. Finally, the enhanced infrared image is output by multiplying the enhanced illumination component and the denoised reflection component. The experimental results show that compared with the traditional Retinex algorithm and the original RetinexNet algorithm, this algorithm can effectively improve the contrast, enrich the image details and texture, and improve the quality of infrared images.
出处
《计算机科学与应用》
2022年第12期2795-2803,共9页
Computer Science and Application