Automatically detecting the canonical orientation of images based on their visual content is an essential part of many computer vision and image processing pipeline. It allows to disregard meta-data such as EXIF which may not be consistently stored or manipulated across different devices, applications, and file formats. Similar to human perception, convolutional neural networks can exploit implicit object recognition and other useful semantic cues to predict the correct orientation of an image. In this work, we leverage the properties of convolutional neural networks and adapt a pre-trained model to the image orientation detection task. It is shown by extensive evaluation that our method can work well on a range of datasets, including relatively low quality photos generated by a wide range of consumer devices. On public benchmarks, our method compares favourably with the state-of-the-art and achieves accuracy very close to that of humans.
Automatic detection of canonical image orientation by convolutional neural networks / Morra, Lia; Famouri, Sina; Cagri Karakus, Huseyin; Lamberti, Fabrizio. - STAMPA. - (2019), pp. 1-6. (Intervento presentato al convegno IEEE 23RD International Symposium on Consumer Technologies (ISCT 2019) tenutosi a Ancona, Italy nel June 19-21, 2019) [10.1109/ISCE.2019.8901005].
Automatic detection of canonical image orientation by convolutional neural networks
Lia Morra;Sina Famouri;Fabrizio Lamberti
2019
Abstract
Automatically detecting the canonical orientation of images based on their visual content is an essential part of many computer vision and image processing pipeline. It allows to disregard meta-data such as EXIF which may not be consistently stored or manipulated across different devices, applications, and file formats. Similar to human perception, convolutional neural networks can exploit implicit object recognition and other useful semantic cues to predict the correct orientation of an image. In this work, we leverage the properties of convolutional neural networks and adapt a pre-trained model to the image orientation detection task. It is shown by extensive evaluation that our method can work well on a range of datasets, including relatively low quality photos generated by a wide range of consumer devices. On public benchmarks, our method compares favourably with the state-of-the-art and achieves accuracy very close to that of humans.Pubblicazioni consigliate
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/11583/2734323