Search

Due to the diversity of text size in scene images, the current box-based methods employ a large amount of fixed-size anchors with different scales to match texts, thus leading to high computational cost. In this paper, we propose to learn the scales of texts and adjust the sizes of anchors accordingly, which can largely reduce the numbers of anchors and therefore significantly reduces the time cost. Moreover, compared to discrete scales used in previous methods, the learned scales are continuous and more reliable. Additionally, we propose Anchor convolution to exploit scaled feature for each anchor by dynamically adjusting the sizes of receptive fields according to the learned scales. Experimental results show that the proposed method significantly improves the computational efficiency of box-based framework(reduce the running time from 0.73s to 0.28s) and enhances its robustness against small texts, while achieving competitive performance with other methods.

For blind image deblurring problem, regularization term met-hod is effective and efficient. Many existing approaches usually rely on carefully designed regularization terms and handcrafted parameter tuning to obtain satisfactory solution. It is complex and difficult. In this paper, we proposed a novel learning-based blind deconvolution method. We learn a Multi-Scale Shrinkage Fields model (MSSF). At each scale, we obtain the nonlinear functions and parameters through the data-driven way. Our method achieved strong robustness against others. It was evaluated on several widely-used natural image deblurring benchmarks, and achieved competitive results.

For blind image deblurring, a good prior knowledge can guide the maximum a posterior (MAP) based algorithms to be away from the trivial solution. Therefore, many existing methods focus on designing effective priors to constrain the solution space. However, blind deconvolution with fixed priors is not robust. And many priors are extremely costly to design and compute. In this paper, we proposed a blind deconvolution method with adaptive priors under the MAP framework. Specifically, we carry out our algorithm under the multi-scale, and at each scale we add specific sparse regularization to standard deblurring formulation. By tunning both the priors and the weights we can give more flexible sparse regularization constraint. After iteration, our algorithm output both latent image and estimated blur kernel, simultaneously. We prove the convergence of the proposed algorithm. Extensive experiments show the effectiveness of our proposed approach.