Image Orientation Estimation And Feature Learning Based On Deep Learning

With the advent of the era of multimedia information,there are a variety of media data in our real life,including images,video,and voice.These media data,especially image data,occupy an increasingly important place in our lives.The widespread use of intelligent terminals such as smartphones and digital cameras has brought about ex-plosive growth of image data,and it has become very important to effectively acquire,analyze,and understand images.In order to analyze and understand the image for the computer,we need to characterize the image.The characteristic expression of the image is to use reasonable algorithm design to obtain distinguishing features in the image.This thesis focuses on the topic about feature representation of the image and dis-cusses the orientation estimation and feature learning of the image in depth.In the feature representation of the image,the rotation transformation of the image is very common and does not affect the semantics of the image.Therefore,it is important to be able to accurately estimate the orientation of the image.We study the orientation estimation of images in both global level and local level.For an image of large-scale outdoor scene,it has a well-defined orientation,so this image is suitable for global level research.For multi-objective images,it has no clearly defined orientation,so this image are suitable for local level research.To address the above two problems,we propose two methods which are elaborated in the following,respectively.Firstly,we propose a global orientation estimation method to align the large-scale outdoor scene images for convenience of the following visual analysis.In the task,we use the neural network to predict the orientation of the outdoor scene image for rotating invariant expression.We use the manually labeled orientation as a reference orientation to learn the orientation of the image.We compared the performance of the orientation estimation of outdoor scene images using a learning-based approach and a traditional manual design approach.Convolutional neural network has a huge advan-tage in our mission,and its performance is significantly better than traditional manual design methods.At the same time,we also compared the performance of the AlexNet model,the MobileNet model,and the VGGNet model when extracting features related to image orientation.MobileNet model uses fewer resources and has fewer errors,so it is more suitable for our task.Secondly,we propose a deep learning method for local patch based orientation estimation as well as feature representation.In the task,We first use the dominant ori-entation of the SIFT feature as the reference orientation for network pre-training,and then use the Siamese architecture to fine-tune parameters.We compare the performance of the SIFT feature dominant orientation estimation method with our method in the fea-ture point matching task.Our method has a certain improvement relative to the domi-nant orientation estimation method of SIFT features.We also compared the changes in experimental results before and after parameters fine-tuning.The use of Siamese archi-tecture networks for parameter fine-tuning has improved performance to some extent.At the same time,we also compared the performance of the SIFT feature expression method and the learning-based method in feature point matching tasks.Convolutional neural network can extract rich and effective image features,so the learning-based fea-ture expression method have better performance.