Fusion of Ensembled UNET and Ensembled FPN for Semantic Segmentation

Image segmentation is an annotation method used to gain a deeper understanding of the images. Semantic segmentation involves constructing a pixel-by-pixel mask of an image by training a neural network. The accuracy of the semantic segmentation algorithms can be improved by eliminating background noise, and computational efficiency can be improved by using the pre-trained networks. This paper proposes a new architecture that ensemble inceptionV3, DenseNet, Resnet34 in the encoder part of UNET and ensemble inceptionV3, Resnet34, and VGG16 in the encoder part of FPN. The ensemble results are fused based on the weighted average and the predictions of the pixels are made on the fused features to perform semantic segmentation. The proposed architecture is implemented on Oxford-IIIT Pet Dataset, created by the visual geometry group, and on the SD saliency 900 dataset. The F1 score, IOU score, and Loss are used to evaluate segmentation model results. The results of the study show that the proposed architecture formed by the fusion of ensembled architectures is more accurate and efficient in segmenting oxford-IIIT pet dataset with the IoU score of 98.68% and segmenting the SD saliency 900 dataset with the IoU score of 66.78%.