一种基于卷积神经网络的区域调光技术

A Convolutional Neural Network Based Local Dimming Technology

由于光线串扰,像素补偿算法难以根据提取出的背光信息进行准确补偿,同时,单一补偿曲线难以适应具有不同亮度特点的图像内容,导致补偿图像的平均质量不高.为了提高像素补偿算法对复杂图像内容的适应性,本文引入神经网络中的编码和解码思想,通过编码网络提取图像深层特征,在解码网络中利用浅层特征的信息对深层特征进行解码,提出了一种联合分类回归的液晶像素补偿神经网络.实验结果表明,此网络得到的像素补偿图像不仅可以提高图像的主观质量,还在对比度、峰值信噪比等客观指标上取得了较好的效果.

Due to the light spreading problem,pixel compensation algorithms are difficult to accurately compensate brightness according to the extracted backlight information.Besides,a single compensation curve is difficult to adapt to the complex image content,resulting in unsatisfactory image quality.In order to improve the adaptability of pixel compensation,the idea of encoding and decoding in neural network was introduced.The deep feature of image was extracted by encoding network,and it was decoded by using the information of shallow feature in decoding network.A kind of classification-regression compensation neural network(CRCNN)was proposed.The experimental results show that the pixel compensation image obtained by this network can not only improve the subjective quality of the image,but also achieve good results in contrast,peak signal-to-noise ratio and other objective indicators.