Overview:

Automatic segmentation and classification of image regions is relevant in various areas of cognitive computing which involves both natural and medical images. While semantic segmentation of regions/objects in  natural images is critical to applications involving augmented reality, automated driving, region/object retrieval etc., automatic segmentation of medical images is crucial for tasks like tumour detection or lesion detection. Segmentation, like any other machine learning task requires large number of training data which involves tedious labelling. Few-shot segmentation techniques involve efforts directed towards learning to segment specific regions/objects in images with limited label data just like humans learning to de-lineate regions with minimal supervision. In this work we propose to innovate techniques that involve multiple levels of feature correlation inside an episodic training framework involving a support set (with segmentation labels) and query images. The novel methods for optimizing multi-level feature correlation computation will be explored as this is imperative to perform the episodic training and inference in realistic time-schedules.  The proposed method will be benchmarked against PASCAL – 5i  and COCO-20i  few-shot segmentation datasets.