Image segmentation; Smooth loss function; Soft argmax

We focus on semantic image segmentation with the usage of deep neural networks and give emphasis on the loss functions used for training the networks. Considering region-based losses, Dice loss, and Tversky loss, we propose two independent modules that easily modify the loss functions to take into account the relationship between the class predictions and increase the slope of the gradient. The first module expresses the ambiguity between classes and the second module utilizes a differentiable soft argmax function. Each of the modules is used before the standard loss is computed and remains untouched. In the benchmark, we involved two neural network architectures with two different backbones, selected two loss functions, and examined separately two scenarios for softmax and sigmoid top activation functions. In the experiment, we demonstrate the usefulness of our modules by improving the IOU and F1 coefficients on the test dataset for all scenarios tested. Moreover, the usage of the modules decreases overfitting. The proposed modules are easy to integrate into existing solutions and add near-zero computational overhead.