A Study On Convolutional Neural Networks For Image Super-resolution And Pedestrian Detection

Deep learning has been developed for decades.Computer vision community has been deeply influenced by powerful learning ability of the hierarchically connected layer architectures,especially convolutional neural networks(CNN).Features,namely locally connected convo-lutions,similar dimension with signals of vision,high performance of non-linear represen-tation and easy implementation on parallel computing devices,of CNN make it felicitously suitable for computer vision tasks,e.g.low-level image processing,mid-level image recog-nition and high-level image semantic understanding.Networks for image super-resolution built by fully convolutional layers could be considered as a multi-layer architecture that projects the input image into convolutional features which have same size of input image.Classification is a very important field of machine learning.For vision,image classification is widely studied by researchers and it has very many applications,e.g.handwriting digits recognition in banking industry.Different from image classification,object detection from image is a higher-level vision task which should recognize the category of objects and locate them.Recently,a series of works,including R-CNN variants,YOLO,SSD,transformed de-tection task into classification or prediction task.This made people leave the sliding window technique which searches the whole area violently and born with long run time.Dilated convolutions support expanding receptive fields without parameter exploration or resolution loss,which turn out to be suitable for pixel-level prediction problems.In chapter 3,we propose Dilated Convolutional networks for image SR,named as DCSR.We analysis the receptive field of dilated convolutions and provide an insight why dilated convolution works for SR.Mixtures of standard and dilated convolutions are employed as building bricks of DCSR.Based on the mixed convolutional layers,we further present a mixed residual block(MR-block)to achieve faster convergence.Moreover,we jointly learn the maps with different scales from a low resolution image to its high resolution one in a single network.Experimental results demonstrate that the proposed method outperforms the state-of-the-art ones in terms of PSNR and SSIM.The improvements of visual effect in large scaling factors are also obvious.Since pedestrians in videos have a wide range of appearances such as body poses,occlusions,and complex backgrounds,pedestrian detection is a challengeable task.In chapter 4,we propose part-level fully convolutional networks(FCN)for pedestrian detection.Based on the detection proposals by a "pre-detector",we adopt deep learning to deal with the proposal shifting problem in pedestrian detection.First,we combine CNN and FCN to align bounding boxes for pedestrians.Then,we perform part-level pedestrian detection based on CNN to recall the lost body parts.Experimental results demonstrate that the proposed method achieves 6.83%performance improvement in log-average miss rate over CifarNet.