Omni-Dimensional Adaptation for MobileNetV3 using Bayesian Hyperparameter Tuning

This paper proposes enhancing MobileNetV3 with Omni-Dimensional Dynamic Convolution (OD-Conv) toovercome CNNs’ limitation of static convolution kernels. OD-Conv introduces multi-dimensional attention to adjust convolution kernels across spatial, input channel, output channel, and number of kernels dimensions, improving feature representation. Bayesian Optimization optimizes hyperparameters efficiently. Experiments show Omni-MobileNetV3 outperforms MobileNetV3 on CIFAR-100, Tiny ImageNet, and medical image datasets, achieving up to 3% accuracy gain while maintaining efficiency. This dynamic convolution method combined with Bayesian tuning achieves state-of-the-art results in image classification