Supervised Deep Learning for Ideal Identification of Image Retargeting Techniques

Image retargeting is the practice of adjusting an image’s dimensions so that its content and visual quality remain preserved as the image is shrunk to fit onto different screens or devices while retaining all its important visual elements and details. Various techniques have been developed for this, such as cropping (CR), Scaling (SCL), seam carving (SC), warping (WARP), Scale-and-Stretch (SNS), Multi-Operator (MULTI), and Shift-Map (SM). However, the best method to retarget a specific image with a specific dimension need to be determined automatically. Thus, this paper proposes deep learning models based on transfer learning, such as Resnet18, DenseNet121, and InceptionV3, to predict the suitable retargeting method for the input image with a specific resolution. We use a dataset of 46,716 images of different resolutions from various retargeting techniques belonging to six categories. Experiments are also conducted with models where the category is fed as a third input and with the resolutions encoded as an extra channel in the image. Furthermore, the models are evaluated with various evaluation metrics. The outcomes demonstrated the effectiveness of the proposed approach for selecting the proper retargeting technique with a best case F1 score of 90%.