本文为奥地利维也纳科技大学（作者：Georg Sperl）的硕士论文，共86页。

大规模目标识别挑战，如ImageNet大规模视觉目标识别挑战或Microsoft上下文中的公共目标挑战，表明卷积神经网络在目标检测和图像分类等计算机视觉问题上取得了最新的性能。卷积神经网络受益于数十万幅图像的数据集，这些数据集覆盖了更多的类内变化，有助于学习鲁棒和不变特征。然而，这些数据集被设计用于一般的物体识别，并且没有类似尺寸的数据集存在于人物识别中。因此，从30多个数据集中收集数据，用于人的检测、分类、分割和跟踪，形成一个用于人物识别的数据源池。提出了一种从该库中提取特定应用数据并训练卷积神经网络进行二值人物分类的方法。此外，分析了非人类对子类标记的性能改进，得到了2.82%的错误率。结果表明，使用我们的人物识别数据集作为人分类任务的预训练集，训练集只有几千幅图像，可以提高8%以上的准确率，总的准确率超过99%。我们的数据集质量通过额外的评估得到了证明。此外，研究结果强调卷积神经网络架构选择的复杂性，并指示在初始化和求解算法的子类标记训练中增加鲁棒性。

Large-scale object recognition challenges such as the ImageNetLarge Scale Visual Object Recognition Challenge or the Microsoft Common Objectsin Context challenge have shown that convolutional neural networks achievestate-of-the-art performance on computer vision problems like object detectionand image classification. Convolutional neural networks benefit from datasetsof hundreds of thousands of images, which cover more intraclass variabilitiesand aid in learning robust and invariant features. However, these datasets aredesigned for general object recognition and no dataset of similar dimensionsexist for person recognition. Therefore, data is collected from over 30datasets for person detection, classification, segmentation and tracking toform a pool of data sources for person recognition. A method of extractingapplication-specific data from this pool and training a convolutional neuralnetwork for binary person classification is proposed. Additionally, performanceimprovements of subclass labeling are analyzed for the nonperson class and anerror rate of 2.82% is achieved. Results demonstrate that using our personrecognition dataset as a pre-training set for person classification tasks withtraining sets of only up to a few thousand images leads to an increase inaccuracy of over 8% to a total accuracy of over 99%. The quality of our datasetis demonstrated by additional evaluation. Furthermore, results emphasize thecomplexity of convolutional neural network architecture choice and indicateincreased robustness in training with subclass labeling with regards toinitialization and solver algorithms.

1. 引言

2. 卷积神经网络

3. 人物识别

4. 相关工作

5. 研究方法

6. 具体实现

7. 评估与结果

8. 结论

更多精彩文章请关注公众号：