Research On Image Recognition And Semantic Segmentation Based On Multi-Stream Convolutional Network

With the rapid development over these years,deep learning based methods have made great breakthrough in realms of artificial intelligence application areas,such as object detection,speech recognition and machine translation.In computer vision community specifically,unlike the traditional methods relying on feature engineering by hand,convolutional neural networks(CNN)based approaches as the representative of deep learning,which have been making breakthrough in almost all subfields of computer vision,have pushed the research frontier forward.In this thesis,we will focus on two important research topics in computer vision based on convolutional neural network: image recognition and semantic segmentation.The image recognition,or rather image classification,is the most fundamental mission in computer vision research.CNN-based methods for image make hierarchical feature extracting and then categorize the image.The core of image recognition is the feature learning process,which is also the basis of all computer vision tasks,so arguably image recognition is the cornerstone of other computer vision missions.Designing a powerful CNN architecture for representation learning is the key point of tackling visual tasks.Meanwhile,semantic segmentation,another fundamental topic of computer vision,is crucial to achieve image understanding,image editing/retrieval and environment perception of robot.Since semantic segmentation is dense pixel categorization,it can be seen as a high-level,more fine classification problem.This thesis first analyzes the existed image classification approaches,and then we propose a multi-stream CNN architecture which deploys a discriminator,and introduce the corresponding training policy adapted from adversarial learning.We demonstrate its efficacy on several benchmark datasets and conduct visualization experiment to further validate the effectiveness of the method.Later,we make extra experiments to diagnosis the discriminator in the method,and propose a novel semantic segmentation approach based on an attention model,which adopts multi-scales path and multi-dilation convolution policy.The method surpasses related works in two benchmark datasets.