基于整数小波变换的指纹信息嵌入方案研究

提出量化指纹编码的思想.将指纹数据转化为二值图像嵌入文件,使攻击者难以发现指纹信息,避免了通过比对编码进行的共谋攻击;设计了整数小波变换嵌入算法,并在实验中得到较好的效果.     

School timetabling for quality student and teacher schedules

The school timetabling problem, although less complicated than its counterpart for the university, still provides a ground for interesting and innovative approaches that promise solutions of high quality. In this work, a Shift Assignment Problem is solved first and work shifts are assigned to teachers. In the sequel, the actual Timetabling Problem is solved while the optimal shift assignments that resulted from the previous problem help in defining the values for the cost coefficients in the objective function. Both problems are modelled using Integer Programming and by this combined approach we succeed in modelling all operational and practical rules that the Hellenic secondary educational system imposes. The resulting timetables are conflict free, complete, fully compact and well balanced for the students. They also handle simultaneous, collaborative and parallel teaching as well as blocks of consecutive lectures for certain courses. In addition, they are highly compact for the teachers, satisfy the teachers’ preferences at a high degree, and assign core courses towards the beginning of each day. Dr. Birbas is currently the Director for Primary and Secondary Education in the Region of Western Greece.     

隐蔽图像后门攻击

目的 图像后门攻击是一种经典的对抗性攻击形式,后门攻击使被攻击的深度模型在正常情况下表现良好,而当隐藏的后门被预设的触发器激活时就会出现恶性结果。现有的后门攻击开始转向为有毒样本分配干净标签或在有毒数据中隐蔽触发器以对抗人类检查,但这些方法在视觉监督下很难同时具备这两种安全特性,并且它们的触发器可以很容易地通过统计分析检测出来。因此,提出了一种隐蔽有效的图像后门攻击方法。方法 首先通过信息隐藏技术隐蔽图像后门触发,使标签正确的中毒图像样本(标签不可感知性)和相应的干净图像样本看起来几乎相同(图像不可感知性)。其次,设计了一种全新的后门攻击范式,其中毒的源图像类别同时也是目标类。提出的后门攻击方法不仅视觉上是隐蔽的,同时能抵御经典的后门防御方法(统计不可感知性)。结果 为了验证方法的有效性与隐蔽性,在ImageN et、MNIST、CIFAR-10数据集上与其他3种方法进行了对比实验。实验结果表明,在3个数据集上,原始干净样本分类准确率下降均不到1%,中毒样本分类准确率都超过94%,并具备最好的图像视觉效果。另外,验证了所提出的触发器临时注入的任意图像样本都可以发起有效的后门攻击。结论 ...     

CentroidNet：轻量快速的乳腺癌Ki67细胞核中心点检测模型

目的 Ki67分数是乳腺癌预后评估的重要指标,计算该分数的关键步骤是检测阴性与阳性癌细胞核。人工检测面临疲劳与主观差异的问题。卷积神经网络有望实现高质量、自动化的细胞核检测,然而需要病理专家为其标注细胞核。为了减轻标注的工作量,不少研究者提出以中心点标注训练卷积神经网络。然而这些方法采用过于复杂的卷积神经网络和后处理流程,未能充分提高易用性和效率、发挥卷积神经网络的质量。对此,提出CentroidNet模型,旨在提高中心点检测的质量、效率和易用性。方法 CentroidNet模型在图像上放置均匀排布的锚点,为每个锚点预测一个候选点,一部分候选点通过基于阈值的筛选策略成为预测点。本文提出最近锚点匹配策略用于生成训练标签,既保证了端到端推理,又规避了其他一对一标签匹配算法所具有的标签抖动问题。本文建议锚点间距应尽可能接近训练集答案点间最短距离的第一百分位数,并指出这样的锚点间距能够在前景标签数、坐标回归难度与效率之间取得良好的平衡。本文在设计卷积神经网络的结构时,没有采纳广为使用的U-Net或特征金字塔（feature pyramid network, FPN）中的多级上采样与旁路连接,反...     

Integrated mathematical optimisation approach for the tower crane hook routing problem to satisfy material demand requests on-site

Optimising decisions around the location and operation of tower cranes can improve the workflow in construction projects. Traditionally, the location and allocation problems involved in tower crane operations in the literature have been solved separately from the assignment of material supply points to demand points and the scheduling of the crane’s activity sequence across supply and demand points on a construction site. To address the gap, this paper proposes a binary integer programming problem, where location of the tower crane, allocation of supply points to material-demanding regions, and routing of hook of the crane based on activity sequencing of the hook across supply and material-demanding regions on site are optimised. The novelty in this work is in the way the crane’s activity scheduling is modelled via mathematical programming, based on routing the hook movement to meet material demand, through minimising tower crane operating costs. A realistic case study is solved to assess the validity of the model. The model is contrasted with results obtained from other solving algorithms commonly adopted in the literature, along with a solution proposed by an experienced practitioner. Results indicate that all instances can be solved when compared to other meta-heuristics that fail to achieve an optimum solution. Compared to the solution proposed by the practitioner, the results of the proposed model achieve a 46% improvement in objective function value. Planners should optimise decisions related to the location of the crane, the crane’s hook movement to meet service requests, and supply points’ locations and assignment to material-demanding regions simultaneously for effective crane operations.     

基于标签噪声鲁棒学习的疾病风险预测

疾病风险预测能够筛查易患人群, 并在早期进行预防干预措施以降低疾病的发生率及死亡率. 随着机器学习技术的快速发展, 基于机器学习的疾病风险预测得到了广泛应用. 然而, 机器学习十分依赖于高质量的标注信息, 医疗数据中存在的标签噪声会给构建高性能的疾病风险预测算法带来严峻挑战. 针对这一问题, 本文提出了一种基于深度神经网络和动态截断损失函数的噪声鲁棒学习方法用于疾病风险预测. 该方法引入动态截断损失函数, 融合了传统交叉熵函数的隐式加权特性和均方差损失函数的标签噪声鲁棒性; 通过构造训练损失下界, 并引入样本动态加权机制减小可疑样本的梯度, 限制可能的带噪样本在训练过程中的权重, 进一步增强模型的鲁棒性. 以脑卒中筛查数据集为例进行实验, 结果表明本文算法在各个标签噪声比例下均能取得良好的预测性能, 可降低疾病风险预测中标签噪声的负面影响, 实现了带有标签噪声数据的鲁棒学习.     

Label distribution learning for scene text detection

Recently, segmentation-based scene text detection has drawn a wide research interest due to its flexibility in describing scene text instance of arbitrary shapes such as curved texts. However, existing methods usually need complex post-processing stages to process ambiguous labels, i.e., the labels of the pixels near the text boundary, which may belong to the text or background. In this paper, we present a framework for segmentation-based scene text detection by learning from ambiguous labels. We use the label distribution learning method to process the label ambiguity of text annotation, which achieves a good performance without using additional post-processing stage. Experiments on benchmark datasets demonstrate that our method produces better results than state-of-the-art methods for segmentation-based scene text detection.     

用于视网膜血管分割的半监督深度学习框架北大核心CSCD

针对目前视网膜血管分割任务中伪标签质量参差不齐,获得高质量的伪标签需要经过筛选的问题,本文提出了一种新的用于视网膜血管分割的半监督深度学习框架。该框架采用分而治之的思想来处理数据,针对有标签数据,采用传统的深度学习方法;针对无标签数据,采用Mean teacher模型,通过对比同一输入的不同形态输出,让模型学习无标签数据之间的共同特征,避免了采用伪标签技术带来的筛选过程。本文将U型网络(u-neural networks,U-Net)、Dense-Net和Ladder-Net 3个基准网络放入该框架,在DRIVE和CHASEDB1数据集上进行训练测试,均取得了较好的分割效果,表明本文框架具有提高网络区分不同阈值像素的能力。     

Exact algorithms for a generalization of the order acceptance and scheduling problem in a single-machine environment

This paper studies a generalization of the order acceptance and scheduling problem in a single-machine environment where a pool consisting of firm planned orders as well as potential orders is available from which an over-demanded company can select. The capacity available for processing the accepted orders is limited and each order is characterized by a known processing time, delivery date, revenue and a weight representing a penalty per unit-time delay beyond the delivery date. We prove that the existence of a constant-factor approximation algorithm for this problem is unlikely. We propose two linear formulations that are solved using an IP solver and we devise two exact branch-and-bound procedures able to solve instances with up to 50 jobs within reasonable CPU times. We compare the efficiency and quality of the results obtained using the different solution approaches.