Multiclass Semantic Segmentation of Mediterranean Food Images.

With the continuous increase of artificial intelligence applications in modern life, the segmentation of images is one of the fundamental tasks in computer vision. Image segmentation is the key for many applications and is backed by a large amount of research, including medical image analysis, healthcare services and autonomous vehicles. In this study we present a semantic segmentation model for food images, suitable for healthcare systems and applications as major part of the dietary monitoring pipeline, trained on an annotation dataset of Mediterranean cuisine food images. To segment the images, we use for feature extraction the ResNet-101 CNN model pre-trained on the ImageNet LSVRC-2012 dataset as a backbone network and the Pyramid Scene Parsing Network - PSPNet architecture for food image segmentation. For the evaluation metric we use the Intersection over Union, where the proposed model achieves a meanIoU score 0.758 in 50 classes of the Mediterranean Greek Food image dataset and 0.933 IoU score in food/non-food segmentation. To evaluate the proposed segmentation model, we train and evaluate a U-Net segmentation model on the same dataset, which achieves meanIoU 0.654 and IoU score 0.901 in multiclass and food/non-food segmentation, respectively.