Dragonflies segmentation with U-Net based on cascaded ResNeXt cells

In cooperation with biologists, we discuss the problem of animal species protection with the usage of modern technologies, namely mobile phones. In our work, we consider the problem of dragonfly image classification, where the aim is given to a preprocessing—segmentation of a dragonfly body from a background. To solve the task, we improve U-Net architecture by ResNeXt cells firstly. Further, we focus on the reasonability of features in neural networks with cardinality dimension and propose the cascaded way of re-using the features among blocks in particular cardinal dimensions. The reuse of the already trained features leads to composing more robust features and more efficient usage of neural network parameters. We test our cascaded cells together with three various U-Net versions for four different settings of hyperparameters with the conclusion that the system of cascaded features leads to higher accuracy than the other versions with the same number of parameters. Also, the cascaded cells are more robust to overfitting the dataset. The obtained results are confirmed on two additional public datasets.