Earticle

An effective visible light and infrared image fusion method using a deep learning framework is designed to obtain a fused image which contains all the features from infrared and visible images. First, the source images are decomposed into low frequency and high frequency sub bands using wavelet transform. Then the low frequency is fused by maximum fusion rule. For the high frequency sub bands a deep learning network is used to find activity level measurements and then fused using the maximum fusion rule. For reconstruction, the optimized orthogonal matching pursuit algorithm and inverse wavelet transform are used.

본 논문에서는 시각적인 자극을 희소 분포로 표현함으로써 입력 자극이 가지고 있는 형태적인 시맨틱을 기억하고 리 콜하는 모델을 제안한다. 사람의 두뇌에서 대뇌 신피질은 정보 처리와 기억을 담당하며 대뇌 신피질에서 뉴런의 동 작 메커니즘은 희소 분포 표현(SDR; Sparse Distributed Representation)으로 모델링 할 수 있다. 이와 같은 희소 분포 표현 기반의 기억 모델은 입력 자극을 기억하는 과정과 유사한 기억을 리콜하는 과정으로 구분하여 고려 할 수 있다. 먼저, 기억 과정은 시각 자극을 희소 분포 표현으로 변환하는 과정이며, 이 과정에서 기본적으로 입력 정보의 형태적인 시맨틱이 유지된다고 생각할 수 있다. 다음으로 이와 같이 기억된 정보와 새로운 입력 자극에 대한 희소 분포 표현들을 비교함으로써 유사한 형태적 시맨틱을 가지고 있는 기억을 리콜할 수도 있다. 본 논문에서는 리 콜 과정에서 형태적 시맨틱의 유사도를 측정하기 위한 기준으로 희소 분포 표현의 중첩률을 사용한다. MNIST (Modified National Institute of Standards and Technology) 데이터를 활용한 제안 모델의 실험 결과 10,000 개의 영상을 기억에 사용하고 주어진 시각 자극에 대한 리콜 실험을 한 결과 희소 분포 표현이 시각 자극의 형태적 시맨틱을 잘 유지하고 있음을 확인하였다.

This paper proposes a model to remember and recall morphological semantic of input visual stimuli using SDR(sparse distributed representation). Neocortex in human brain is in charge of information processing and memory, and operation of neurons in neocortex can be modeled by sparse distributed representation. This memory model based on sparse distributed representation can be considered as memory precess and recall process separately. First, memory process means a process to convert input visual stimuli to sparse distributed representation, and in this process, morphological semantic of input visual stimuli can be preserved. Next, recall process can be considered by comparing sparse distributed representation of new input visual stimulus and remembered sparse distributed representations. Superposition of sparse distributed representation is used to measure similarities. Experimental results using 10,000 images in MNIST(Modified National Institute of Standards and Technology) data set show that the sparse distributed representation of the proposed model efficiently keeps morphological semantic of the input visual stimuli.

Recently, an emerging method called image sparse representation has attracted more attentions. The method has been proved to be effective in various image processing applications. It is important to note that few sparse representation methods fail to analyze the aliasing in satellite image restoration. To address the problem, firstly, we employ adaptive reciprocal cell as a image quality estimation tool, which can analyze the satellite image degradation factors including aliasing, blur and noise. Then, with the help of the powerful tool, the estimation about the satellite image quality is introduced into the sparse representation model . Experiment results show that our method can produce good quality restored results.

We propose a novel image inpainting approach in which the exemplar and the sparse representation are combined together skillfully. In the process of image inpainting, often there will be such a situation: although the sum of squared differences (SSD) of exemplar patch is the smallest among all the candidate patches, there may be a noticeable visual discontinuity in the recovered image when using the exemplar patch to replace the target patch. In this case, we cleverly use the sparse representation of image over a redundant dictionary to recover the target patch, instead of using the exemplar patch to replace it, so that we can promptly prevent the occurrence and accumulation of errors, and obtain satisfied results. Experiments on a number of real and synthetic images demonstrate the effectiveness of proposed algorithm, and the recovered images can better meet the requirements of human vision.

In this paper, a machine-learning approach called Sparse Representation-based Classification (SRC) is used for automatic chord recognition in music signals. We extracted different Pitch Class Profile (PCP) features from raw audio and achieved sparse representation of classes via 1 -norm minimization on feature space to recognize 24 major and minor triads. This recognition model is evaluated on MIREX’09 dataset including the Beatles corpus. Our method is compared with various methods that entered the Music Information Retrieval Evaluation eXchange (MIREX) in 2014 towards the audio chord estimation of MIREX’09 dataset in Audio Chord Estimation task of MIREX. Experimental results demonstrate that our method has good accuracy rate in recognizing maj-min chords.

In this paper, we have proposed an optimized sparse representation algorithm based on Log-Gabor (Sparse Representation-based Classification Based on Log-Gabor, Log-GSRC), which applies local features information of samples to the sparse representation method. Actually, SRC (Sparse Representation-based Classification) is using a linear correlation between the samples of one class which can be assumed that these samples exist in a subspace, and also can be linear represented with each other. It is a global representation and it ignores the local features information of the samples, while in the case of there are a smaller number of training samples per class, SRC will obtain an inaccurate classification result which may correspond to one and more classes in the process of sparse decomposition. However, the Log-GSRC combines global and local features information of the samples and also improves the robustness of SRC. The experimental results clearly showed that Log-GSRC has much better performance than SRC and also has much higher recognition rates than SRC in face recognition.

Image denoising is a basic problem in image processing, due to the image structure has the characteristic of self-similar, using the ideas of nonlocal, this paper proposes a non-local denoising method based on sparse representation, the structure information of image is improved after denoising, at the same time making similar image tiles have similar sparse representation, image reconstruction effect is better, through the numerical simulation the results show that the method has good application value.

The emergence of a large number of databases for capturing 3D human motions has made the efficient analysis and processing of human motion data to effectively use these databases a new challenge. To reduce high-dimension complexity, a dimensional feature based on the 3D spatial–temporal characteristic should be extracted from human motions. Moreover, the motion data should be re-expressed by sparse representation to realize the projection from high dimensional data to a low-dimensional subspace. The different motions should then be recognized and classified to obtain the automatic recognition and automatic retrieval of 3D human motions.

In this paper, a machine-learning approach called Sparse Representation Classification(SRC) Viterbi Algorithm is proposed for automatic chord recognition in music. We extracted Pitch Class Profile(PCP) features or Log PCP from raw audio and achieved sparse representation of classes via -norm minimization on feature space to recognize 24 major and minor triads. This recognition model is evaluated MIREX'09 dataset including the Beatles corpus. Our method is also compared with various methods that entered the Music Information Retrieval Evaluation exchange (MIREX) in 2013 and 2014. Experimental results demonstrate that our method has good accuracy rate in recognizing signal chord and has fewer train data.

In this paper, we propose a fast and robust tracking method based on reversed sparse representation. Be different from other sparse representation based visual tracking methods, the target template is sparsely represented by the candidate particles which are gotten by particle filter. In order to improve the robustness of the method, we use a target template set. Meanwhile, a two level competition mechanism is also introduced. In the first level, each target template is sparsely represented and all the candidate particles compete with each other by a similarity calculation, which is based on sparse coefficients. Then, the winners construct a target candidate set. In the second level, all the target candidates in the target candidate set compete with each other and the one which is the most similar to the template set is considered as the target. In addition, a template set update strategy is proposed to adapt the appearance variations of the target. Experimental results on challenging benchmark video sequences demonstrate that the proposed tracking algorithm performs favorably against several state-of-the-art methods.

In this paper, we present a new approach to reconstruct a high resolution (HR) image from a low resolution (LR) input image based on a two dimensional (2D) sparse method. The new method consists of three phases. Firstly, the nonlinear feature of the input LR image is divided into the linear subspace, and then LR-HR dictionaries are learned to reduce the blurred artifacts of the image. Secondly, 2D sparse representation and self-similarity are developed to strengthen and enhance the image structure. Finally, the final HR image is achieved by reconstruction of all HR patches. Simulation results demonstrated that our proposed method achieved superior results on real images, and shows various improvements in terms of PSNR and SSIM values as compared with some other competent methods.

In this paper, we propose a robust tracking method with a novel appearance model based on fragments-based PCA sparse representation. It samples non-overlapped local image patches within the templates in PCA subspace. Then, the candidate local image patches are sparse represented by the local template patches in PCA subspace. Finally, tracking is continued using the particle filter for propagating sample distributions over time. In addition, the templates are updated online based on incremental subspace learning .Using the fragments-based PCA templates rather than the image templates facilitates the tracker to handle significant illumination and pose change as well as occlusion. Experimental results on challenging videos show that our method can track accurately and robustly, and outperform many other state-of-the- art trackers.

Automatic brain tumor segmentation from multispectral magnetic resonance imaging (MRI) data is an important but a challenging task because of the high diversity in the appearance of tumor tissues among different patients and in many cases similarity with the normal tissues. In this paper, we propose a fully automatic technique for brain tumor segmentation from multispectral human brain MRIs. We first use the intensities of different patches in multispectral MRIs to represent the features of both normal and abnormal tissues and generate a dictionary for following tissue classification. Then, the sparse representation classification (SRC) is applied to classify the brain tumor and normal brain tissue in the whole image. At last, the Markov random field (MRF) regularization introduces spatial constraints to the SRC to take into account the pair-wise homogeneity in terms of classification labels and multispectral voxel intensities. Our method was evaluated on 20 multi-modality patient datasets with competitive segmentation results.

To make sure human face can be more accurately identified in various poses, an identification method based on the characteristics human face image have been proposed in the thesis. First, it is a design for the optimal sampling matrix to acquire compressed measurement data; then it adopts iterative method based on the close loops of l2 and l1 in normal form to get an estimate on human face images, since the method based on norm l2 can measure the relevancy of human face images in the space and time of the continuous time point, and the method based on norm l1 mainly uses the modified total variation method and basis pursuit noise-reduction method. The simulation design has suggested that the method adopted in the thesis can achieve a higher human face identification rate and a remarkable promotion effect.

Superresolution image reconstruction technique uses single or a series of low-resolution images to reconstruct a high resolution image without changing the hardware devices, while improving image quality and the spatial resolution of the image. High resolution means the image with a higher pixel density, can provide more details. In this paper, a novel image superresolution algorithm based on sparse representation is studied. During over-complete dictionary of the training phase, the proposed method improves two aspects including feature extraction and dimension reduction. In the feature extraction process, combining the second derivative with the gradient direction, we construct a new descent direction to improve gradient method. The convergence speed of the new algorithm is faster than the gradient method and can get better results. Then improved two-dimensional Principal Component Analysis (2DPCA) algorithm is used to reduce the dimension, it could eliminate the correlation of the image lines and column. Experiment results show that this method of image reconstruction is better and faster for high resolution image reconstruction.

Facial recognition (FR) is a challenging area of research due to difficulties with robust FR when the number of training samples is very small. The state-of-the-art sparse representation-based classification (SRC) shows very excellent FR performance. However, the recognition rate of SRC will drop dramatically when the number of training samples per class is very limited. To solve these issues, we propose a weighted multi-classifier optimization and sparse representation based (WMSRC) method for FR, which efficiently combines the local and global characteristics of face images. A face image is firstly divided into continuous but non-overlapped blocks by multi-resolution based blocking and each block is sparsely represented over the corresponding set of blocks of all training samples. The multi-scale SRC classifiers are then established and associated with different weights based on sub-block dictionary learning. According to the multiple voting results of the classifiers, the weights of multi-classifiers are optimized by a least-squares optimization equation with 2l-norm regularization. Finally, the classification results of all the blocks are combined by a weighted fusion criterion. Our experiments show that the WMSRC algorithm outperforms many existing block-based sparse representation classification algorithms, especially for FR when the available training samples per subject are very limited.

In this paper, we propose a novel algorithm to deal with the problem of visual tracking in some challenging situations, which is based on sparse representation and multi-scale block. To build target templates, we select distinguishable features between the target and background in each frame of video sequences, dictionary is built by the multi-scale block of target templates. Then, particle filter generates filter distribution in the next frame, the moving target is framed by affine transformation. To describe the current state of the target, we calculate posterior probability for each particle. Finally, the templates are updated online. The experimental results show that the proposed algorithm is superior in accuracy than the classical tracking algorithm, and it has better robustness against to the target posture changes, partial occlusion and illumination variations.

Sparse representation has been successfully applied to visual tracking to find the target with the minimum reconstruction error from the target templates subspace. Traditional sparsity-based trackers handle corruptions and occlusions of the observation by introducing a set of trivial templates. However, the performance is not so satisfactory in practice. It is because the trivial templates unable to model heavy occlusions effectively, and the likelihood computation and the template update processes do not take full advantage of the occlusion information. In this paper, we propose a novel tracking method taking advantage of local sparse representation to detect occlusions during the tracking sequence. In our method, the target is divided into local patches. We analyze the spatial distribution of the samples employed by the local sparse representation, and determine the occlusion state for each patch respectively. The occluded patches are disregard, only the unoccluded ones are considered for reconstruction and likelihood computation. In addition, a dynamic template update strategy with occlusion handling is introduced to alleviate the drift problem. Experiments on challenging video sequences demonstrate that the proposed tracking algorithm performs favorably against several state-of-the-art methods.