基于阶段注意力机制的电力线提取算法

电力系统维护是电力系统稳定运行的重要保障，应用智能算法的无人机电力巡检则为电力系统维护提供便捷。电力线提取是自主电力巡检以及保障飞行器低空飞行安全的关键技术，结合深度学习理论进行电力线提取是电力巡检的重要突破点。本文将深度学习方法用于电力线提取任务，结合电力线图像特点嵌入改进的图像输入策略和注意力模块，提出一种基于阶段注意力机制的电力线提取模型（SA-Unet）。本文提出的SA-Unet模型编码阶段采用阶段输入融合策略（Stage input fusion strategy， SIFS），充分利用图像的多尺度信息减少空间位置信息丢失。解码阶段通过嵌入阶段注意力模块（Stage attention module，SAM）聚焦电力线特征，从大量信息中快速筛选出高价值信息。实验结果表明，该方法在复杂背景的多场景中具有良好的性能。     

A Tradeoff Between Accuracy and Speed for K-Means Seed Determination

With a sharp increase in the information volume, analyzing and retrieving this vast data volume is much more essential than ever. One of the main techniques that would be beneficial in this regard is called the Clustering method. Clustering aims to classify objects so that all objects within a cluster have similar features while other objects in different clusters are as distinct as possible. One of the most widely used clustering algorithms with the well and approved performance in different applications is the k-means algorithm. The main problem of the k-means algorithm is its performance which can be directly affected by the selection in the primary clusters. Lack of attention to this crucial issue has consequences such as creating empty clusters and decreasing the convergence time. Besides, the selection of appropriate initial seeds can reduce the cluster’s inconsistency. In this paper, we present a new method to determine the initial seeds of the k-mean algorithm to improve the accuracy and decrease the number of iterations of the algorithm. For this purpose, a new method is proposed considering the average distance between objects to determine the initial seeds. Our method attempts to provide a proper tradeoff between the accuracy and speed of the clustering algorithm. The experimental results showed that our proposed approach outperforms the Chithra with 1.7% and 2.1% in terms of clustering accuracy for Wine and Abalone detection data, respectively. Furthermore, achieved results indicate that comparing with the Reverse Nearest Neighbor (RNN) search approach, the proposed method has a higher convergence speed.     

变分自编码器模型综述

变分自编码器（VAE）作为深度隐空间生成模型的一种，近年来其表现性能取得了极大的成功，尤其是在图像生成方面。变分自编码器模型作为无监督式特征学习的重要工具之一，可以通过学习隐编码空间与数据生成空间的特征映射，进而在输出端重构生成输入数据。梳理了传统变分自编码器模型及其衍生变体模型的发展与研究现状，并就此做了总结和对比，最后分析了变分自编码器模型存在的问题与挑战，并就可能的发展趋势做了展望。     

结合位图切割和区域合并的彩色图像分割

就经典分水岭图像分割算法中存在的过分割问题，提出一种结合位图切割和区域合并的彩色图像分割算法。对原始彩色图像通过空域梯度算子求其梯度图像，并利用位图切割重建梯度图像；对新梯度图像进行分水岭预分割；对预分割图像基于异质性最小原则进行区域合并，并获得最终分割结果。相比于现有的同类方法，该算法引入位图切割，抑制噪声对分割结果的影响，在边缘模糊处分割准确，得到符合人类视觉的较小分割区域数目，同时在运行效率上提高。     

正则化图形模糊聚类及鲁棒分割算法

针对现有图形模糊聚类算法合理性差和抗噪能力弱的问题，提出嵌入对称正则项的图形模糊聚类鲁棒算法。将样本聚类所对应的中立度与拒分度相结合构造对称正则项，嵌入现有图形模糊聚类所对应的目标函数；同时，利用像素邻域所对应的均值信息辅助当前像素聚类并构造了空间信息约束正则项，采用拉格朗日乘子法获得正则化图形模糊聚类鲁棒分割算法。不同噪声干扰图像分割结果表明，所建议的分割算法是有效的，相比现有的鲁棒模糊聚类分割算法具有更强的抑制噪声能力。     

Learning a Deep Predictive Coding Network for a Semi-Supervised 3D-Hand Pose Estimation

In this paper we present a CNN based approach for a real time 3D-hand pose estimation from the depth sequence. Prior discriminative approaches have achieved remarkable success but are facing two main challenges: Firstly, the methods are fully supervised hence require large numbers of annotated training data to extract the dynamic information from a hand representation. Secondly, unreliable hand detectors based on strong assumptions or a weak detector which often fail in several situations like complex environment and multiple hands. In contrast to these methods, this paper presents an approach that can be considered as semi-supervised by performing predictive coding of image sequences of hand poses in order to capture latent features underlying a given image without supervision. The hand is modelled using a novel latent tree dependency model (LDTM) which transforms internal joint location to an explicit representation. Then the modeled hand topology is integrated with the pose estimator using data dependent method to jointly learn latent variables of the posterior pose appearance and the pose configuration respectively. Finally, an unsupervised error term which is a part of the recurrent architecture ensures smooth estimations of the final pose. Experiments on three challenging public datasets, ICVL, MSRA, and NYU demonstrate the significant performance of the proposed method which is comparable or better than state-of-the-art approaches.      

自适应天牛须优化与K均值聚类的图像分割算法研究

在机器识别中,图像分割是重要的一个步骤,传统分割手段存在一定缺陷。针对传统K均值聚类分割的初始聚类中心敏感的缺陷进行了优化,利用自适应天牛须优化算法,避免了这一问题。通过实验结果表明,该算法(ABASK)对图像进行分割,既可以保证图像轮廓的分割,同时还可以更多地保留图像细节。     

CLOI-NET: Class segmentation of industrial facilities’ point cloud datasets

Shape segmentation from point cloud data is a core step of the digital twinning process for industrial facilities. However, it is also a very labor intensive step, which counteracts the perceived value of the resulting model. The state-of-the-art method for automating cylinder detection can detect cylinders with 62% precision and 70% recall, while other shapes must then be segmented manually and shape segmentation is not achieved. This performance is promising, but it is far from drastically eliminating the manual labor cost. We argue that the use of class segmentation deep learning algorithms has the theoretical potential to perform better in terms of per point accuracy and less manual segmentation time needed. However, such algorithms could not be used so far due to the lack of a pre-trained dataset of laser scanned industrial shapes as well as the lack of appropriate geometric features in order to learn these shapes. In this paper, we tackle both problems in three steps. First, we parse the industrial point cloud through a novel class segmentation solution (CLOI-NET) that consists of an optimized PointNET++ based deep learning network and post-processing algorithms that enforce stronger contextual relationships per point. We then allow the user to choose the optimal manual annotation of a test facility by means of active learning to further improve the results. We achieve the first step by clustering points in meaningful spatial 3D windows based on their location. Then, we apply a class segmentation deep network, and output a probability distribution of all label categories per point and improve the predicted labels by enforcing post-processing rules. We finally optimize the results by finding the optimal amount of data to be used for training experiments. We validate our method on the largest richly annotated dataset of the most important to model industrial shapes (CLOI) and yield 82% average accuracy per point, 95.6% average AUC among all classes and estimated 70% labor hour savings in class segmentation. This proves that it is the first to automatically segment industrial point cloud shapes with no prior knowledge at commercially viable performance and is the foundation for efficient industrial shape modeling in cluttered point clouds.     

Food recognition using neural network classifier and multiple hypotheses image segmentation

ABSTRACT

This paper proposes the multiple-hypotheses image segmentation and feed-forward neural network classifier for food recognition to improve the performance. Initially, the food or meal image is given as input. Then, the segmentation is applied to identify the regions, where a particular food item is located using salient region detection, multi-scale segmentation, and fast rejection. Then, the features of every food item are extracted by the global feature and local feature extraction. After the features are obtained, the classification is performed for each segmented region using a feed-forward neural network model. Finally, the calorie value is computed with the aid of (i) food volume and (ii) calorie and nutrition measure based on mass value. The experimental results and performance evaluation are validated. The outcome of the proposed method attains 0.947 for Macro Average Accuracy (MAA) and 0.959 for Standard Accuracy (SA), which provides better classification performance.