自适应融合局部和全局特征的图像质量评价

Adaptive Fusion of Local and Global Features for Image Quality Assessment

无参考图像质量评价通过算法来量化图像质量的失真程度。有效建立失真位置与周围空间的依赖关系能提高质量预测的准确性,但目前基于卷积神经网络的无参考图像质量评价方法,仅通过传统的卷积对局部失真区域进行特征提取,无法有效地获取全局的失真关系,容易弱化对失真扭曲等特征表示。因此,提出了一种基于自适应融合局部和全局特征的图像质量评价算法。在待评价图像上进行特征提取时,自适应地构建围绕每个空间位置的长距离空间和通道间的依赖关系,通过全局失真关系来增强局部特征信息的表征能力;增强图像的细节信息,并在不同尺度的特征层上自适应地融合局部和全局失真信息,整合更加丰富的失真特性,进而提高特征的判别性;再将多个尺度上的不同失真信息进行融合获得最终的质量评价得分,这种融合可以避免图像浅层信息的损失。为验证模型的有效性,在真实失真和合成失真数据集上进行实验对比分析,结果表明,在真实失真数据集LIVEC上SROCC达到0.867,对图像质量的预测更符合人类对质量的感知。

No-reference image quality assessment aims to quantify image quality the degree of visual distortion of image quality by algorithms.Establishing the dependency between distortion location and surrounding space effectively is able to improve the accuracy of quality prediction,but the current convolutional neural network based no-reference image quality evaluation method only extracts features from local distortion regions by traditional convolution,which cannot effectively obtain the global distortion relationship and tends to weaken the feature representation of distortion,etc.Therefore,an image quality evaluation algorithm based on adaptive fusion of local and global features is proposed.First,feature extraction on the image to be evaluated is performed by adaptively constructing long-range spatial and inter-channel dependencies around each spatial location to enhance the characterization of local feature information through global distortion relationships.Then,detail information of the image is enhanced and local and global distortion information is adaptively fused on feature layers at different scales to integrate richer distortion characteristics and thus improve the discriminatory nature of the features.Finally,the different distortion information at multiple levels is fused to compute the final quality evaluation score,which can avoid information loss in lower layers.To verify the effectiveness of the model,experimental comparative analysis is carried out on the real distortion and synthetic distortion datasets,and the results show that the SROCC reaches 0.867 on the real distortion dataset LIVEC,and the prediction of image quality is more in line with human perception of quality.