Search

This paper presents an approach for image cartoonization. By observing the cartoon painting behavior and consulting artists, we propose to separately identify three white-box representations from images: the surface representation that contains smooth surface of cartoon images, the structure representation that refers to the sparse color-blocks and flatten global content in the celluloid style workflow, and the texture representation that reflects high-frequency texture, contours and details in cartoon images. A Generative Adversarial Network (GAN) framework is used to learn the extracted representations and to cartoonize images. The learning objectives of our method are separately based on each extracted representations, making our framework controllable and adjustable. This enables our approach to meet artists' requirements in different styles and diverse use cases. Qualitative comparisons and quantitative analyses, as well as user studies, have been conducted to validate the effectiveness of this approach, and our method outperforms previous methods in all comparisons. Finally, the ablation study demonstrates the influence of each component in our framework.

Ultrasonography is an important tool and has been widely used in clinical applications, however, the physicians and surgeons still often suffers great difficulties in diagnosis and treatment due to the high speckle noise of ultrasound images. Existing speckle reduction methods usually over-smooth low contrast features since they are sensitive to contrast variations in the images. In our paper, we propose a novel learning-based scheme to effectively remove the speckle noise while preserving the image features. This is achieved by incorporating a denoising algorithm — non-local simultaneous sparse coding (NLSSC) which is modified by adding a pre-clustering process guided by feature asymmetry (FA). The proposed pre-clustering procedures can help better train the dictionaries in NLSSC and make them more feature-aware. Qualitative and quantitative experiments were carried out to evaluate the performance of our methods against some state-of-the-art on both synthetic and clinical ultrasound images. Experimental results demonstrate that our approach has effective performance and can produce good speckle reduced images.

Intrinsic image decomposition is an important topic in computer vision and computer graphics applications. However, this is a challenging problem by adopting the information of a single image. Therefore, additional priors or supplementary information such as multiply images or user interactions are necessary to address this problem. In this paper, we propose a novel scheme to use multiple images for intrinsic image decomposition, based on a similar strategy in robust principal component analysis (RPCA). RPCA utilizes the fact that the reflectance layer is common in multiple lights of a scene, and attempts to decompose the data matrix constructed from input images into a low-rank matrix and a sparse matrix. This is possible if the sparse matrix is sufficiently sparse, which is often not the case in computer vision applications. Moreover, the weighting parameter between the low-rank and sparse matrices greatly affects the accuracy of the results, and tuning this parameter can be tricky. This paper proposes a rank-constrained PCA algorithm (RCPCA) for solving background recovering problems. Fixing the rank of the low-rank matrix to be 1 allows RCPCA to better recover the low-rank matrix from the data matrix. Comprehensive tests show that RCPCA produces more stable and accurate results than RP-CA.

In this paper, we proposed a novel edge-based method for intrinsic image decomposition. To be specific, to address this ill-posed problem via single image, we employed the information of chromaticity gradient to guide recovering the reflectance gradient map, and regularized the Retinex assumption by the l 0 -norm to render the reflectance layer piece-wise smooth. Our method is simple but effective. Qualitative together with quantitative experiments were carried out on some public datasets. The comparison results show that our method runs much faster than the state-of-art methods with comparable performance.

Accurate segmentation of brain tumor from MR image is crucial for the diagnosis and treatment of brain cancer. We propose a novel automated brain tumor segmentation method based on a probabilistic model combining sparse coding and Markov random field (MRF). We formulate the brain tumor segmentation task as a pixel-wise labeling problem with regard to three classes: tumor, edema and healthy issue. For each class, dictionary learning is performed independently on multi-modality gray scale patches. Sparse representation is then extracted based on a joint dictionary which is constructed by combing the three independent dictionaries. Finally, we build the probabilistic model aiming to estimate maximum a posterior (MAP) probability by introducing the sparse representation into likelihood probability and prior probability using the Markov random field (MRF) assumption. Compared with traditional methods, which employed hand-crafted low level features to construct the probabilistic model, our model can better represent the characteristics of a pixel and its relation with neighbors based on the sparse coefficients obtained from the learned dictionary. We validated our method on the MICAAI 2012 BRATS challenge brain MRI dataset and achieved comparable or better results compared with state-of the-art methods.