Exploring An Easy Way for Imbalanced Data Sets in Semantic Image Segmentation

Abstract

In recent years, deep convolutional neural networks have gradually become the preferred method for image processing. After the development of Classification, Detection and Segmentation, a large variety of state-of-the-art models and algorithms have emerged in the field. However, for some specific data sets or tasks, not all methods are applicable, which is inconvenient to researchers. This paper took the data set provided in the airbus ship detection challenge in Kaggle as an example to explore an easy and effective method for segmentation tasks of data sets with class imbalance. This paper used U-Net with a pre-trained ResNets model, and tried different methods to explore the feature of the set. In the process of training ResNets, this paper proposed a new convolutional block structure which is inspired by Fibonacci sequence, but the effect is not good. In the end, the mF2 values of the models this paper trained achieved good results, which is better than the model of the combined training of ResNets and ordinary U-Net34. Moreover, the training parameters are less than that. This paper believe that this simple and effective training method will bring convenience to researchers in related fields.