Few-shot learning with background subtraction

Few-shot learning for image classification aims to classify the image by only using few images as supporting samples. In the past several years, few-shot learning has achieved a huge improvement in image classification. In the recent work, such as meta-transferlearning (MTL) and Few-shot Adaptive Faster R-CNN have achieved a higher accuracy. In this paper, we are trying to combine three different methods together which are YOLOV2 model, Mask RCNN and our fewshot learning model. When a CNN wants to recognize animals in photos, there is a huge chance that even features that are supposed to represent trees will be encoded as belonging to those animals. Our main idea is to using YOLO algorithm, Mask RCNN and Opencv functions to reduce the noise and background as much as possible and keep our main object as it is. We would like to train our model using the image that only contain the object itself. We show that this approach is helpful to improve the accuracy in few-shot learning image classification.