基于深度CNN的改进弱监督学习方法设计与验证

针对包含目标、尺度和平移变化较强的空间信息难以获取大量训练样本的问题，提出一种基于深度卷积神经网络(deep convolutional neural network，DCCN)的弱监督学习方法，从3个层面对当前卷积神经网络进行扩展。为了提取分辨率更高的局部特征，同时考虑到全卷积网络(full convolution network，FCN)在全监督式学习下的高效性能，使用FCN作为后端模块；为了获取更多的通用特征，增加一个多映射弱监督学习的传输层，对与补充性类模态相关的多个局部特征进行显式学习；为了优化训练过程，改进了池化层，使用全局图像标签进行训练，将空间得分聚合为全局预测。使用图像分类、弱监督逐点目标定位和图像分割3种常用的机器视觉任务进行评估。多个公开数据库的实验结果表明，所提方法能够有效地学习强局部特征，具有良好的分类和定位效果。

Design and verification of improved weakly supervised learning method based on deep CNN

Aiming at the problem that it is very hard to obtain a large number of training samples using spatial information containing multiple-targets, scale change and strong shift changing, a deep convolutional neural network (DCCN) method based on weakly supervised learning is proposed, in which the convolutional neural network is extended from three aspects. Firstly, In order to extract local features with higher resolution, and taking the efficiency of full convolution network (FCN) under full supervised learning into account, the adoption of FCN is being as a back-end module. Secondly, to get more general features, a transport layer of multiple-mapping weakly supervised learning is adopted. Multiple local features related to complementary class modality are explicitly learned. Thirdly, to optimize the training process, pool layer is improved and global image labels are used for training, and the spatial score is aggregated into global prediction. Three kinds of machine vision recognition tasks are used to evaluate: image classification, weakly supervised point by point target location and image segmentation. The experimental results in several open databases show that the proposed method can learn strong local features, and have good classification and location effect.