Enhancing Transfer Learning Network Classification Accuracy on Degraded Images by Network Structural Modification

Transfer learning approach is widely employed in various applications. However, this approach on image classification has its challenges, one of them is that its performance would deteriorate when training image quality is not perfect. In this study, structural modifications on the transfer learning network are applied to mitigate the negative impact caused by image degeneration. We modify the structure via several methods, namely fine-tuning the last layers of the pre-trained model, adding fully-connected layers, and reducing the depth of the pre-trained model. Upon examining the impact of these modifications when common image corruption occurs, we have found that both fine-tuning and reducing the depth of the network improves performance while adding fully-connected layers does not lead to a significantly positive impact. These empirical results have demonstrated to what extent structural modifications can compensate for the transfer learning network performance deterioration due to image corruption. The results help to further broaden the scope of the transfer learning application to degraded image classification.