Research On Small Object Segmentation In Images

Image segmentation is a popular field in the computer vision,and plays a key role in many situations such as autonomous driving,geographic information systems and medical image analysis.Genarally,small objects are unignorable in these situations such as pedestrians and sign symbols in autonomous driving scenarios,roads in remote sensing images,diseased areas and cells in medical image.However it is difficult to segment small objects by existing algorithms,because they have fewer pixels,less obvious features,and relatively complex backgrounds.Traditional algorithms mainly segment images by extracting features like colors,shapes and positions,leading to poor performance in complex backgrounds and bad generalization performance.With the development of computer hardware and machine learning,convolutional neural networks based image segmentation models have greatly improved the segmentation accuracy compared with traditional algorithms.However,it is still difficult to segment small objects in complex backgrounds.This paper studies and improves the image segmentation algorithm based on convolutional neural network.The main contributions of this thesis are as follows:1.In the thesis,Deeplab v3+,a convolutional neural network based semantic segmentation model is studied,and a local enhancement module is added to recalculate the spatial extent of the object by learning the spatial operator,therefore highlighting the detailed features such as edges and small objects in high-level features.This thesis introduces a metric based on shared boundaries between classes aiming at improving the loss function and introduce an encoder network to estimate the metric.The encoder network is then appended to the segmentation model for end-to-end training.Experiments show that the revised semantic segmentation model has greatly improved the segmentation of small objects.2.Mask R-CNN,one of efficient instance segmentation models,is investigated and experiments are conducted on the cell nucleus dataset.In order to improve the segmentation accurancy of small objects,we add a bottom-up branch to the feature pyramid network which shortens the information path between low-level features and high-level features.Thus the localization information in low-level features can improve the low detection of small objects and also improve the average accuracy of the model.Then,a fully connected conditional random field model is appended to the model.The color information and position information of the pixels are fully considered,so that the segmentation results are more refined.