Partial VGG16: reformed network for classification under partial occlusion
Bohang Sun, Bo Chen, Yuandan Deng · 2023
Standard deep Convolutional Neural Networks (CNNs), even those with encoder-decoder transformers, often struggle to generalize well when faced with partial occlusion in images. To address this issue, we were inspired by the success of inpainting models that are particularly sensitive to the missed areas in an image. Therefore, we propose to replace the conventional convolutional layers in CNNs with partial convolution blocks, which are commonly used in image inpainting and other related tasks. This new structure allows CNNs to focus only on the valid parts of an image by ignoring the masked parts. To evaluate this approach, we conducted experiments on a model enhanced with partial convolution blocks, by artificially removing rectangular areas from the original images. The results indicate that this model is more sensitive to the non-occluded parts of the object and achieves higher precision than the standard VGG16 network, with a 5% (±2%) increase in top-1 accuracy, has reached to 94.11% in top-2 accuracy.