Research On Image And Feature Enhancement Methods For Low-resolution Object Recognition

The image classification(recognition)task is a classic computer vision task.The development of computer vision and even deep learning have benefited from the indepth research on this task.Image classification technology also plays an important role in daily life,e.g.,social media,medical imaging,security monitoring,remote sensing,etc..Due to device performance limitations,complex imaging environments,and long imaging distances,images used for classification will inevitably experience degradation,resulting in low-resolution images.Low-resolution degradation will lead to the loss of information in the image,which will affect the accuracy of image classification.To improve the accuracy of low-resolution image classification,it is necessary to reconstruct the lost information in the image.However,the existing reconstruction methods usually face two problems:(1)The number of parameters and computation of the reconstruction models are too large,which leads to the high cost of practical application.(2)The reconstruction algorithm relies on semantic labels for guidance,and the algorithm fails when there are no semantic labels for training.To address above problems,this thesis conducts research in two aspects:For the first problem,this thesis proposes a joint training method based on a bidirectional mapping network to achieve a balance between speed and accuracy.Inspired by the fact that the transformation between low-resolution and high-resolution images is an inverse operation,we design a lightweight bidirectional mapping network in this chapter.The network decomposes low-resolution images into high-frequency and low-frequency information and uses a reversible neural network to enhance these two types of information.In addition,the network provides auxiliary information for lowresolution image enhancement by learning the inverse task.Finally,the computational complexity is further reduced by introducing a sampling module.The experimental results show that the method proposed in this thesis can achieve a classification accuracy similar to other methods with fewer parameters.For the second problem,this thesis proposes a low-resolution image classification method based on long-distance feature dependency capture to get rid of the dependence on semantic labels.This thesis analyzes the distribution of low-resolution image features with different structures in the depth representation space through statistical experiments and reveals that the degradation of low resolution will lead to image patches with different structures having similar feature distributions in the depth representation space.Therefore,a feature enhancement network based on the self-attention mechanism is proposed.The network captures long-range feature dependencies through a self-attention mechanism for disambiguating local degraded features.In this thesis,a pre-trained classification network is used to extract the features of high-resolution images,which are used to train the supervision signal of the feature enhancement network,thereby eliminating the dependence on semantic labels.The experimental results show that the method proposed in this thesis can significantly improve the classification accuracy of low-resolution datasets.