Handling Data Imbalance for Improving Blurriness Estimation using Convolutional Transformer
Hyun-Yong Lee, Nac‐Woo Kim, Jungi Lee, Seok‐Kap Ko · 2023
Image deblurring is important pre-processing for various computer vision tasks. In this paper, as one approach for improving image deblurring, we are interested in blurriness estimation. For this, we first propose a model for blurriness estimation. Adopting the convolutional Transformer, we try to extract meaningful features from a blurry image to be used for blurriness estimation. Then, using the proposed model, we examine the usefulness of the known techniques for handling data imbalance issue, that is widely observed in real-world scenarios. Through the experiments using the RealBlur dataset, we show that the weighted loss is not effective in solving the data imbalance issue. On the contrary, the oversampling technique is useful, particularly in improving the estimation performance of the rare data while slightly sacrificing the estimation performance of the prevalent data.