Abstract: For the difficulty of achieving high-precision classification in deep learning algorithms for small training samples, a small-sample classification algorithm based on reinforcement learning and relational network is proposed. An automatic cropping model based on reinforcement learning aesthetic aware is adopted in the process of image preprocessing, and an optimal cropped image is output by constructing an aesthetic aware reward function so as to retain the most characteristic part of the image. The relation network architecture is utilized, the training samples and test samples in the auto-cropped small-sample image are subjected to extract feature through an embedded module in the relation network. The extracted features are cascaded to the feature map, and the cascaded feature maps are fed into the relational module in the relational network for comparison, and the resulting relation scores in the range of 0 to 1 are used as comparison results to determine the category to which the test samples belong. Experiments on small-sample datasets and comparison with the existing methods, experiments show that this method can achieve high precision of small-sample classification.

Key words: small-sample classification, reinforcement learning, automatic cropping, relational network, feature maps

摘要： 针对少量训练样本在深度学习算法中难以实现高精度分类的问题，提出一种融合强化学习和关系网络的小样本分类算法。采用图像预处理过程中基于强化学习的美学意识图像自动裁剪模型，通过构建美学意识奖励函数来输出最佳裁剪图像，从而保留图像最具特征部分。利用关系网络模型，将自动裁剪后的小样本图像中的训练样本图像与测试图像通过关系网络中的嵌入模块进行特征提取。将提取后的特征进行特征映射级联，并将级联后的特征映射馈送到关系网络中的关系模块中进行比较，将最终产生的0到1范围内的关系评分作为比较结果，从而判断测试图像所属的类别。在小样本数据集上进行实验并与现有方法进行对比，实验表明该方法能够实现较高精度的小样本分类。

关键词: 小样本分类, 强化学习, 自动裁剪, 关系网络, 特征映射