摘要
为了有效地测量不同类型的失真图像的质量,提出一种结合Curvelet(曲波)变换和神经网络的无参考图像质量评价算法(Curvelet IQA)。首先对失真图像进行曲波变换得到曲波系数,对曲波系数利用非对称广义高斯拟合,并找到曲波方向信息的能量特征值,得到一组特征评价值。同时将BP神经网络进行布谷鸟搜索算法优化,利用优化后的神经网络与曲波域相结合的方法对无参考失真图像进行预测。该算法将得到的特征值送入到神经网络中进行训练,建立特征统计值和主观评价值的非线性映射关系,随后利用映射关系得到输入图像的预测值。对LIVE数据库进行验证,实验结果表明与当前相关文献的方法相比,该算法具有更高的一致性和准确性。该算法没有针对特定的失真类型提取统计特征值,而是提取可以测量多种失真类型的统计特征值。
In order to effectively measure the quality of different types of image distortion,propose a no-reference image quality assess- ment method based on Curvelet transform and neural network,called CurveletIQA. First, get Curvelet coefficient by Curvelet transform for the distorted image, the Curvelet coefficient by asymmetric generalized Gaussian fitting, and finding the energy eigenvalue of wave direc- tion information, get a set of characteristic value. Simultaneously,have a cuckoo search algorithm to optimize the BP neural network and use the method of combination of the optimized neural network and Curvelet domain to predict the distorted image without reference. The resultant feature values into the neural network is trained to establish the nonlinear mapping relationship between statistics and subjective evaluation values. Then use the mapping relationship to obtain the predicted value of the input image. The LIVE database is verified. The experimental results show that compared with the current method of relevant literature, the algorithm can effectively evaluate a variety of distortion types, with higher consistency and accuracy.
出处
《计算机技术与发展》
2015年第7期86-90,共5页
Computer Technology and Development
基金
四川省教育科研重点项目(13ZA0169)
核废物与环境安全国防重点学科实验室项目(13zxnk12)
关键词
无参考图像质量评价
曲波变换
自然图像
非对称广义高斯分布
no-reference image quality assessment
Curvelet transform
natural images
asymmetric generalized Gaussian distribution
