Efficiency analysis of ITN loss function for deep semantic building segmentation

In using deep convolutional neural networks (DCNN) for semantic building segmentation, the standard loss functions such as cross-entropy usually create significant deviations in extracting buildings’ boundaries due to misalignment between the extracted edge map and the ground truth. Motivated to address this limitation, the current investigation presents a novel methodology for measuring boundary shifts by utilizing the concept of inverse transform network (ITN) to define the boundary loss function for capturing the spatial relationships between pixels in an image and the target boundaries. The proposed DCNN model is designed based on encoder-decoder architecture for semantic segmentation, which gradually increases the dimensionality of feature maps by embedding a specific module in the network decoder and making modifications. The proposed segmentation network's efficacy is evaluated by utilizing two distinct building datasets: the WHU aerial building dataset and the Massachusetts dataset. Our implementation results on both datasets demonstrate the efficacy of the proposed approach in generating meticulously structured segmentation outputs, boasting remarkable accuracy, particularly within boundary regions. The proposed method surpasses other particular networks (U-NetPlus, PsPNet and DeepLabV3 +) across all evaluation criteria and achieves values of 0.9513 and 0.8472 for the F1-score in WHU and Massachusetts dataset, respectively. Utilizing the suggested boundary loss function (Loss_ITN) during network training resulted in a notable increase of 0.59 in the IOU index of the WHU dataset compared to the training solely reliant on Loss_CE. Moreover, a remarkable 0.46 improvement was achieved in the IOU index of the Massachusetts dataset.