用激活地图集探索神经网络的算法研究

Algorithmic Research on Exploring Neural Networks with Activation Atlases

摘要: 当今信息时代，人工智能的应用几乎渗透到了各种领域，如航空航天图像处理技术的地质探测，军事目标的侦测制导，医学上的临床病理诊断，如今它已经真正地融入到我们的日常生活当中，虽然机器视觉系统得到了广泛应用，但是要理解它到底是如何“分辨”这些事物，将水果A归类为苹果而把水果B归类为梨，仍然不为我们所知。本文主要以“激活地图集”为核心，首先从网络上收集带有标签的图片形成数据集和使用已有的数据集，利用tensorflow自行搭建一个卷积分类神经网络对两类数据集进行训练，最终从训练模型中得到一个较为准确的分类结果。其次就是地图集的构建，将输入图像运行到神经网络中的某一层，然后经过特征提取能够还原图像，经过图像整合并形成一张地图集。总的来说，它不仅能够从这些图像中识别出其形状、颜色，还能够把这些特征结合起来形成特定的场景或者物体。

Abstract: In this current information age, machine learning technology application has begun to gradually enter our lives, and the image processing technology is also developing rapidly. The machine vision system has been widely used, such as geological detection of aerospace image processing technology, detection and guidance of military targets, and clinical pathological diagnosis in medicine, it has been truly integrated into our daily life. However, when we try to understand how it “recognizes” these things, like categorizing fruit A as an apple and fruit B as a pear, it’s still unknown. This article mainly focuses on “activating atlases”. Firstly, it collects tagged images from the network to form a data set or an existing data set. Then, it uses tensorflow technology to build a convolutional neural network to train the two types of image data sets to get a more accurate classification result from the training model. The second is the construction of the atlas, which runs the input image to a certain layer in the neural network, and then through feature extraction can restore the image and form an atlas. In general, it can not only recognize its shape and color from these images, but also combine these features to form a specific scene or object.