- 09-08-2024
- MACHINE LEARNING
Doshisha University researchers created techniques to enhance image cropping models, preventing accidental removal of critical elements like watermarks.
Researchers at Doshisha University have developed advanced adversarial techniques to improve automatic image cropping models. By introducing imperceptible noisy perturbations, these methods ensure that crucial parts of images, like copyright information, are preserved and not inadvertently cropped.
Their study, published in IEEE Access, features two approaches: a white-box method using internal model gradients and a black-box method employing Bayesian optimization. Both methods enhance cropping accuracy and fairness, addressing biases and legal risks associated with automatic image cropping. This research marks a significant step towards more reliable and transparent AI systems.