Intuitionistic Neuro-Fuzzy Systems for Semantic Image Segmentation

In recent years semantic segmentation of images has grown to be of vital importance in applications from autonomous driving to medical image analysis. Deep learning techniques have significantly enhanced the performance of semantic segmentation compared to traditional approaches. However, semantic segmentation remains a challenging task because of fuzzy boundaries of objects, and the presence of fine-grained detail in images. To address these issues this paper introduces the Adaptive Intuitionistic Neuro-Fuzzy Segmentation System (AINFSS). The CNN backbone of AINFISS is a pre-trained EfficientNetB0 model that extracts feature maps. Intuitionistic fuzzy sets are used in trainable blocks to minimize uncertainty in the feature maps generated by the CNN. The fuzzy blocks minimize the degree of hesitancy, a metric for pixel class uncertainty, that helps to produce better quality feature maps. The feature maps with reduced uncertainty are deconvolved to produce the segmented image. Experimental results on the Oxford-IIIT Pets dataset indicate that AINFSS attains a mean IoU score of 75.61%, 3% higher than state of the art models.