Sharji University in the UAE recently unveiled a groundbreaking solution in the realm of automatic image adaptation for various screens and devices. Their team of experts has devised deep learning models that can accurately predict the optimal sizes of images, taking into consideration factors such as resolution and screen format. These DL models, utilizing transfer training, have proven to be superior to traditional methods of pruning and resizing images.
The development of deep learning architectures like Resnet18, Densenet121, and InceptionV3 has played a key role in the analysis of image structure, object recognition, and classification. According to scientists, this new approach significantly reduces the loss of visual information when resizing images, ensuring high-quality display on all types of devices.
Traditionally, methods for altering images such as pruning, scaling, Seam Carving, and distortion require manual selection of the optimal method, leading to potential distortions or improper cropping, especially on screens with non-standard aspect ratios. The innovative approach developed by Sharji University automates the selection process, determining the best method for adapting images to a given resolution.
For the development of their models, researchers utilized a large dataset consisting of 46,716 images across different resolutions and categories. Through experiments that incorporated additional parameters like category and encoded image resolution, the accuracy of the algorithms reached an impressive 90% by metric F1.
The automation of retargeting method selection has become increasingly crucial with the wide array of devices and screens in use today. Leveraging deep learning not only allows for image analysis but also enables the creation of models capable of considering complex relationships to effectively choose the optimal adaptation method.
The authors highlight the high potential of this new technology for further advancement. Plans for the future include expanding the database and incorporating additional retargeting methods to enhance algorithm accuracy and adaptability to a broader range of tasks. However, details regarding the commercial availability of this solution have yet to be disclosed.