Abstract
The lack of long-range dependence in convolutional neural networks causes weaker performance in generative adversarial networks(GANs) with regard to generating image details. The self-attention generative adversarial network(SAGAN) use the self-attention mechanism to calculate the correlation coefficient between feature vectors, which improves the global coherence of the network. In this paper, we put forward an improved-self-attention GANs(Improved-SAGAN) to improve the method for calculating correlation in the SAGAN. We can better measure the correlation between features by normalizing the feature vectors to eliminate as many errors caused by noise as possible. As the network learns the global information by calculating the correlation coefficient between all features, it can make up for the defects of local receptive field in the convolution network. We replace the conventional one-hot label with multi-label to obtain more supervised information for generative adversarial networks. We generate dairy goat images based on auxiliary condition generative adversarial network(ACGAN) incorporating the normalized self-attention mechanism and prove that images generated under multi-label are of higher quality than images generated under one-hot label. The generative results of different networks on the public dataset are compared by the inception score and FID evaluation algorithms, and we propose a new evaluation algorithm called SSIM-Mean to measure the quality of generated dairy goat images to further verify the effectiveness of the improved-self-attention GANs.
Highlights
Generative adversarial networks have been well applied in many fields in recent years as a kind of unsupervised learning
In 2014, Alec Radford et al combined the GANs with a convolution neural network(CNN) [1]to design a new adversarial network called DCGAN [2], which makes the adversarial learning a major step forward in the image generation
The self-attention mechanism first obtains the correlation between all features on the feature map, and multiplies the correlation coefficient matrix by the input feature map to obtain global information on features, which overcomes the defect of local receptive field in the CNN
Summary
Generative adversarial networks have been well applied in many fields in recent years as a kind of unsupervised learning. Tang: Dairy Goat Image Generation Based on Improved-SAGANs self-attention mechanism to the GANs for the first time. The feature correlation is calculated by the dot product to make the images generated by the SAGAN have global coherence, which contributes to large-resolution images to be more realistic in detail. The feature vectors are normalized before calculating the self-attention matrix by the dot product, which contributes to avoiding calculation errors due to noises. We use Wasserstein distance with the gradient penalty as the loss function and replace the one-hot label with multi-label which includes three types of supervised information (colors, backgrounds, behaviors) to train the networks, and generate dairy goat images controlled by multi-label. Improvement for calculating the degree of feature correlation in the SAGAN by normalizing features before performing the dot product on the feature map. Proposition of the SSIM-Mean algorithm which is based on the SSIM algorithm and is used to measure the quality of generated dairy goat images
Sign-in/Register to view highlights & summary Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
