IEEE Access · 2021 · 11 citations · 16 references
ComputingFood ScienceComputer VisionFeature EngineeringFruit ScienceGenerative ModelingImage ClassificationImage Feature EngineeringImage RegistrationObject RecognitionAesthetic Attribute RecognitionAttribute RecognitionImage-to-image TranslationImage-to-text TranslationNeural Style TransferStyle Transfer For VideoTraining
CycleGAN can realize image translation and style transferring among unpaired images. However, it will easily generate inappropriate image results when the number and shapes of objects in the style offering image and the source image are greatly different. The paper proposed an improved network, named arCycleGAN, which introduced the mechanism of attribute registration into CycleGAN to solve the problem. The arCycleGAN can transfer the freshness styles from the style offering images to the unpaired input source images. The generated target images will have the freshness attributes of the style offering images, while maintaining the shapes and key features of the input source images. The realization of mechanism of attribute registration consists of three modules. The first module is attribute recognition module, which can identify and label the attributes of objects in images. The second module is image pre-screening module, which selects appropriate image subset as screened training set from raw image set according to the attributes of the input source images. The third module is similarity matching module, which matches the images in screened training set based on the similarity. The generator and discriminator in the new network are similar to that in the CycleGAN network. Experimental results demonstrate the effectiveness and better performance of the arCycleGAN. Compared with the CycleGAN, the new network can generate more convincing images. It can generate the target images of similar quality based on a smaller training set and less training time than the original CycleGAN. For generating images of similar quality, the number of images in the required training set can be reduced by 50%, while training time is reduced by 5.8%.
16
Deep Residual Learning for Image Recognition
Kaiming He, Xiangyu Zhang, Shaoqing Ren et al. · 2016 · 222.1K citations · Full text
Gradient-based learning applied to document recognition
Yann LeCun, Léon Bottou, Yoshua Bengio et al. · Proceedings of the IEEE · 1998 · 58.2K citations · Full text
Unpaired Image-to-Image Translation Using Cycle-Consistent Adversarial Networks
Jun-Yan Zhu, Taesung Park, Phillip Isola et al. · 2017 · 21.9K citations · Full text