Fitting objects with implicit polynomials by deep neural network

Implicit polynomials (IPs) are considered as a powerful tool for object curve fitting tasks due to their simplicity and fewer parameters. The traditional linear methods, such as 3L, MinVar, and MinMax, often achieve good performances in fitting simple objects, but usually work poorly or even fail to obtain closed curves of complex object contours. To handle the complex fitting issues, taking the advantages of deep neural networks, we designed a neural network model continuity-sparsity constrained network (CSC-Net) with encoder and decoder structure to learn the coefficients of IPs. Further, the continuity constraint is added to ensure the obtained curves are closed, and the sparseness constraint is added to reduce the spurious zero sets of the fitted curves. The experimental results show that better performances have been obtained on both simple and complex object fitting tasks.     

Building efficient CNN architecture for offline handwritten Chinese character recognition

Deep convolutional neural networks-based methods have brought great breakthrough in image classification, which provides an end-to-end solution for handwritten Chinese character recognition (HCCR) problem through learning discriminative features automatically. Nevertheless, state-of-the-art CNNs appear to incur huge computational cost and require the storage of a large number of parameters especially in fully connected layers, which is difficult to deploy such networks into alternative hardware devices with limited computation capacity. To solve the storage problem, we propose a novel technique called weighted average pooling for reducing the parameters in fully connected layer without loss in accuracy. Besides, we implement a cascaded model in single CNN by adding mid output to complete recognition as early as possible, which reduces average inference time significantly. Experiments are performed on the ICDAR-2013 offline HCCR dataset. It is found that our proposed approach only needs 6.9 ms for classifying a character image on average and achieves the state-of-the-art accuracy of 97.1% while requires only 3.3 MB for storage.     

基于局部区域动态覆盖的3D点云分类方法

局部几何形状的描述能力, 对不规则的点云形状表示是十分重要的. 然而, 现有的网络仍然很难有效地捕捉准确的局部形状信息. 在点云中模拟深度可分离卷积计算方式, 提出一种新型的动态覆盖卷积(dynamic cover convolution, DC-Conv), 以聚合局部特征. DC-Conv的核心是空间覆盖算子(space cover operator, SCOP), 该算子通过在局部区域中构建各向异性的空间几何体覆盖局部特征空间, 以加强局部特征的紧凑性. DC-Conv通过在局部邻域中动态组合多个SCOP, 实现局部形状的捕捉. 其中, SCOP的注意力系数通过数据驱动的方式由点位置自适应地学习得到. 在3D点云形状识别基准数据集ModelNet40, ModelNet10和ScanObjectNN上的实验结果表明, 该方法能有效提高3D点云形状识别的性能和对稀疏点云的鲁棒性. 最后, 也提供了充分的消融实验验证该方法的有效性. 开源代码发布在https://github.com/changshuowang/DC-CNN.     

基于深度确定性策略梯度的粒子群算法

在传统的粒子群优化算法(PSO)中,所有粒子都遵循最初设定的一些参数进行自我探索,这种方案容易导致过早成熟,且易被困于局部最优点。针对以上问题,该文提出了一种基于深度确定性策略梯度的粒子群优化算法(DDPGPSO),通过构造神经网络分别实现了动作函数和动作价值函数,且利用神经网络可以动态地生成算法运行所需要的参数,降低了人工配置算法的难度。实验表明DDPGPSO相比9种同类算法在收敛速度和寻优精度上均有较大的提升。     

Mathematical representation of 2D image boundary contour using fractional implicit polynomial

Implicit polynomial (IP) fitting is an effective method to quickly represent two-dimensional (2D) image boundary contour in the form of mathematical function. Under the same maximum degree, the fractional implicit polynomial (FIP) can express more curve details than IP and has obvious advantages for the representation of complex boundary contours. In existing studies, algebraic distance is mainly used as the fitting objective of the polynomial. Although the time cost is reduced, there are problems of low fitting accuracy and spurious zero set. In this paper, we propose a two-stage neural network with differentiable geometric distance, which uses FIP to achieve mathematical representation, called TSEncoder. In the first stage, the continuity constraint is used to obtain a rough outline of the fitting target. In the second stage, differentiable geometric distance is gradually added to fine-tune the polynomial coefficients to obtain a contour representation with higher accuracy. Experimental results show that TSEncoder can achieve mathematical representation of 2D image boundary contour with high accuracy.