基于序列对抗域适应的智能裁剪算法

智能裁剪任务一直受到缺乏训练数据的困扰,目前还局限于公开数据集中.因为实际应用场景与训练场景之间存在域迁移,文中提出基于序列对抗域适应的智能裁剪算法.首先,通过实验证实裁剪数据集GAICD和CPC之间存在域迁移问题.然后,构造由美学评分模块和对抗域适应模块组成的算法.美学评分模块用于预测图像的美学评分,并辅助提取面向裁剪任务的不变特征.对抗域适应模块实现基于对抗的域适应学习.不同裁剪数据集之间的域迁移实验及室内/室外场景之间的域迁移实验均验证文中算法的有效性.

Listwise Adversarial Domain Adaption Algorithm for Image Cropping

Image cropping is short of training data for its high threshold for annotation. Current research on image cropping is confined on public datasets. Grounded on domain shift between training domain and practical application scene, a listwise adversarial domain adaption algorithm for image cropping is proposed in this paper. Firstly, the domain shift between two image cropping datasets, GAICD and CPC, is proved. Then, an image cropping model composed of an aesthetic evaluation module and an adversarial domain adaptation module is constructed. Aesthetic evaluation module is employed to predict the aesthetic score of current image and assist the model to extract the invariant features for cropping task. Adversarial domain adaptation module is exploited to realize adversarial based domain adaptation learning. Domain migration experiments between different cropping datasets and between different scene domains verify the effectiveness of proposed algorithm.