基于注意力机制的个性化图像美学质量评估

Research on Evaluation of Personalized Image Aesthetic Quality Based on Attention Mechanism

图像美学质量评估在图像推荐和图像编辑美学等行业具有良好的应用前景。大部分美学质量评估方法存在两个问题:第一,只建立了通用美学评估模型,而忽略了个性化审美,这样的评估方法只适用于大众审美的评估,无法区分出不用用户之间的差异;第二,往往只从图像本身特征和整体布局进行评估,没有考虑主体的突出性和位置合理性,由于主体特征对评估结果有较大的影响,以致忽视主体的做法会使得评估结果具有较差的完备性。针对这些问题,文中设计了一种基于注意力机制的个性化图像美学质量评估方法。该方法在个性化残差网络的基础之上引入了注意力机制,既保持了个人在美学质量评估的主观偏好,又将主体的显著性因素加入评分过程。在Flickr数据集上的实验中,该算法的平均SROCC相关系数表现为0.667,相较于单一的图像美学质量评估在Flickr数据集上的表现,性能提升了3%。

Image aesthetic quality assessment has a good application prospect in image recommendation and image editing aesthetics.There are two problems in most aesthetic quality assessment methods.On the one hand,only the general aesthetic assessment model is established,while the individual aesthetic assessment is ignored.Such an assessment method is only applicable to the public aesthetic assessment and cannot distinguish the differences between users.On the other hand,the evaluation is usually conducted only from the features of the image itself and the overall layout,without considering the prominence and rationality of the subject.Since the subject features have a great impact on the evaluation results,the practice of ignoring the subject will make the evaluation results have poor completeness.Aiming at these problems,we design a personalized image aesthetic quality evaluation method based on attention mechanism.This method introduces the attention mechanism on the basis of the personalized residual network,which not only maintains the subjective preference of the individual in the aesthetic quality assessment,but also adds the significance factor of the subject into the scoring process.In experiments on Flickr data sets,the average SROCC correlation coefficient of this algorithm is 0.667,which is 3%better than the performance of a single image aesthetic quality assessment on Flickr data sets.