Researchers on the Kwame Nkrumah College of Science and Expertise (KNUST) have developed a wiser approach for synthetic intelligence (AI) to research high-resolution pictures with out dropping necessary particulars.
Conventional AI fashions shrink pictures to save lots of computing energy, however this typically reduces accuracy. The brand new strategy, known as WaveNet, retains pictures sharp whereas making the method extra environment friendly.
WaveNet makes use of a way known as the wavelet packet remodel (WPT) to interrupt pictures into smaller, significant components earlier than feeding them into AI fashions.
It additionally features a particular characteristic, the wavelet-adaptive environment friendly channel consideration (WAECA) module, which helps the AI give attention to essentially the most helpful components of a picture.
Assessments on standard AI fashions like ResNet-50 and MobileNetV2 confirmed that WaveNet improves accuracy whereas considerably decreasing processing prices.
For instance, when examined on the Caltech-256 dataset, a WaveNet-enhanced ResNet-50 achieved 72.47% accuracy, outperforming the usual model (70.65%) whereas utilizing far fewer computational assets.
By making AI fashions each smarter and quicker, this innovation might assist enhance picture recognition in functions like medical imaging, surveillance, and autonomous autos, with out requiring costly {hardware} upgrades.
This analysis, revealed in Engineering Experiences, was supported by the Deutsche Gesellschaft für Internationale Zusammenarbeit (GIZ) GmbH via the Accountable Synthetic Intelligence Lab at KNUST.
Authors of the research embody: Albert Dede, Henry Nunoo-Mensah, Emmanuel Kofi Akowuah, Kwame Osei Boateng, Prince Ebenezer Adjei, Francisca Adoma Acheampong, Isaac Acquah, Jerry John Kponyo.
DISCLAIMER: The Views, Feedback, Opinions, Contributions and Statements made by Readers and Contributors on this platform don’t essentially signify the views or coverage of Multimedia Group Restricted.
Source link
