Convolutional Neural Networks (CNNs) benefit from fine-grained details in high-resolution images, but these images are not always easily available as data collection can be expensive or time-consuming. Transfer learning pre-trains models on data from a related domain before fine-tuning on the main domain, and is a common strategy to deal with limited data. However, transfer learning requires a similar domain with enough available data to exist, and transferability varies from task to task. To deal with limited high-resolution data we propose resolution transfer: using low-resolution data to improve high-resolution accuracy. For resolution transfer, we use Continuous kernel CNNs (CKCNNs) that can adapt their kernel size to changes in resolution and perform well on unseen resolutions. Training CKCNNs on high-resolution images is currently significantly slower than CNNs. We lower the inference costs of CKCNNs to enable training on high-resolution data. We introduce a CKCNN parameterization that constrains the frequencies of kernels to avoid distortions when the kernel size is changed, improving resolution transfer accuracy. We improve fine-tuning with a High-Frequency Adaptation module that complements our constrained kernels. We demonstrate that CKCNNs with kernel resolution adaptation outperform CNNs for resolution transfer tasks with no fine-tuning or with limited fine-tuning data. We compare to transfer learning, and achieve competitive classification accuracy with an ImageNet pre-trained ResNet-18. Our method provides an alternative to transfer learning that uses low-resolution data to improve classification accuracy when high-resolution data is limited.

Generative adversarial networks (GANs) are a popular method for image generation, capable of state-of-the-art. Despite this popularity, the generation of fabric patterns remains somewhat unexplored. A possible reason for this could be that there is no publicly shared dataset large enough to train models. However, research into this topic could be relevant because designers can take up to multiple weeks for a single design, but a trained GAN could generate thousands of designs per day. Next to this fabric patterns have characteristics that make them different from the datasets commonly used right now, which means an investigation could yield new insights into the capabilities of GANs. This research investigated the performance of several GAN models on the task of fabric pattern generation and compared these to another popular method of image generation: variational autoencoders. By using open-source implementations of GAN models, and a dataset of fabric patterns provided by Vlisco, a Dutch fabric company, the performance of GANs has been evaluated. To protect the interests of Vlisco, this dataset cannot be shared. The results show clear promise in the capabilities of GANs to replicate these complex patterns, but these results are not yet of the same quality that GANs have been able to achieve in other domains. However, these results are an improvement from those of previous research. The results achieved by variational autoencoders are similar to those of GANs, but state-of-the-art GANs perform better. Due to limited time and resources, only low-resolution samples have been generated for evaluation. In the future, higher-resolution images could be generated which could yield additional insights into the current capabilities of GANs in this domain.