Search

We propose Object-oriented Neural Programming (OONP), a framework for semantically parsing documents in specific domains. Basically, OONP reads a document and parses it into a predesigned object-oriented data structure (referred to as ontology in this paper) that reflects the domain-specific semantics of the document. An OONP parser models semantic parsing as a decision process: a neural net-based Reader sequentially goes through the document, and during the process it builds and updates an intermediate ontology to summarize its partial understanding of the text it covers. OONP supports a rich family of operations (both symbolic and differentiable) for composing the ontology, and a big variety of forms (both symbolic and differentiable) for representing the state and the document. An OONP parser can be trained with supervision of different forms and strength, including supervised learning (SL) , reinforcement learning (RL) and hybrid of the two. Our experiments on both synthetic and real-world document parsing tasks have shown that OONP can learn to handle fairly complicated ontology with training data of modest sizes., Comment: accepted by ACL 2018

A pyrene derivative as the donor and a butyl-viologen as the acceptor were used to construct a novel charge-transfer cocrystal consisting of mixed-stacking structure and having switchable photoconductivity stemming from the donor-acceptor heterojunctions within the lattice.

In this paper, we investigate object tracking in video sequences and propose a particle filter based tracking algorithm with color and texture information fusion, in which the target model is jointly represented by spatial-weighted color histogram and LBP (Local Binary Patterns) texture histogram. The property of local grayscale or color invariance for LBP operator makes it more reliable to measure the spatial structure of local image texture. The system is less sensitive to illumination changes and partial occlusions, and can be able to track objects in diverse conditions. Experimental results demonstrate that the performance of the proposed method is more robust and accurate than the original color based method, especially when tracking objects with similar color appearance to the background and partial occlusions.

This paper presents a modified codebook model for real-time moving foreground detection. The proposed method is an effective combination of background modeling and motion detection. Without a long training sequence, the background model can be represented in a compressed form, a series of codebooks, which means sample background values for each pixel are quantized into codebooks that can used in detection process. In this way, we can capture the structural variation of background in different conditions such as periodic-like motion , hostile environment or change of scene caused by moving object over a long period of time under limited memory. Compared with the original codebook model, this proposed method is more efficient in computation and takes up less memory. Experimental results show that the proposed algorithm is effective, quick for motion detection, and can meet the demands of real-time applications.

This paper proposes a novel colorization algorithm for monochrome still images based on fuzzy clustering and distance transformation. Given small amount of typical color scribbles manually marked on the input grayscale image, the followed colorization process consists of four main steps: color scribble extraction, distance transformation and spatial weight estimation, fuzzy clustering and luminance weight estimation, and weighted color blending. By propagating the local color hints from the given scribbles to the whole grayscale image, the final output color image can be produced. Experimental results show that when the content of grayscale image is relatively simple, the proposed colorization method can efficiently gain a good colorized image with only a small number of scribbles. The algorithm is simple and easy to implement. Moreover, it takes only several seconds to complete the colorization process but can get visually vivid result.

This paper presents a simple exemplar-based method for 2D image texture synthesis. The contribution of the proposed algorithm lies in the effective combination of MEBC (Minimum Error Boundary Cut) and LBP (Local Binary Patterns) feature. Given a texture sample, the corresponding LBP image is firstly generated, then block matching is carried out based on LBP image and the candidate block with MEBC boundary is pasted on the output image. Experimental results for a wide range of texture samples demonstrate that the performance of this LBP-based method is equal to or better than that of the original image quilting algorithm, especially for semi-structured and stochastic textures. Furthermore, the proposed method can obtain a particularly effective synthesized result even for the texture samples that image quilting fails. For regular texture synthesis, the calculation of MEBC can be further omitted and therefore the resulted simplified version of the proposed method can even achieve a good performance in a shorter time.