Image Instance Segmentation Methods Based On Attention Mechanism & Super-pixel

As a digital image processing technique,image instance segmentation plays a very important role.It is rapidly developed in these years by the propulsion of deep learning.However,there is still some defects in this technique.In this paper,we mainly discuss two enhancing methods for image instance segmentation.The first approach is the attention based optimization method.The attention mechanism is a data processing approach in machine learning,which is widely used in various different machine learning tasks like speech recognition,image recognition,natural language processing,etc.The attention mechanism can help the algorithm to locate the target object’s key part in a fast way,which leads to processing more efficiently and accurately.The workflow we designed is,during the feature extracting procedure of a certain instance segmentation method,utilizing the attention mechanism,not treating pixels at every location equally but giving more weights to the more important locations;then feeding the features filtered and optimized by attention to the next processing stage to executing detection bounding boxes regression and instance masks generation;and finally obtaining the more comprehensive and accurate segmentation results.The second approach is a simple post-processing method for instance segmentation.Instances segment images into parts with rich semantics while less texture consistency.While super-pixels segment images into parts with great texture consistency while less semantics.So we design a method,joining super-pixel to the instance segmentation workflow,in order to enhance the instance segmentation results.The workflow is,firstly calling a certain instance segmentation method(for example,Mask-RCNN)on the image to get prediction masks preliminary;then utilizing super-pixels as the assistant information to modify the prediction masks;and finally obtaining the better segmentation results.Our method is train-free,while it can refine the instance segmentation masks.Our experiments performed on multiple neural networks and the MSCOCO dataset demonstrate the effectiveness of our method.