<i>Padding Module</i>: Learning the Padding in Deep Neural Networks

During the last decades, many studies have been dedicated to improving performance of neural networks, for example, network architectures, initialization, and activation. However, investigating importance effects learnable padding methods in deep learning remains relatively open. To mitigate gap, this paper proposes a novel trainable Padding Module that can be placed model. The optimize itself without requiring or influencing model’s entire loss function. train itself, constructs ground truth predictor from inputs by leveraging underlying structure input data supervision. As result, learn automatically pad pixels border its images feature maps. contents are realistic extensions simultaneously facilitate downstream task. Experiments shown proposed outperforms state-of-the-art competitors baseline methods. For has 1.49% 0.44% more classification accuracy than zero when tested on VGG16 ResNet50.