膠囊網路中卷積層之效能分析

近年來，卷積神經網路在電腦視覺領域的問題上一直都有著出色的表現。卷積神經網路是由卷積層及池化層及全連接層所組成的，尤其是卷積層在卷積神經網路有著重要的功能，主要是在擷取不同的特徵以便後續做分類。在2017年時由學者Hinton提出的膠囊網路，是透過全新向量概念及動態路由演算法的類神經網路模型，且在膠囊網路中原始架構採用了二層架構的卷積層。要提升膠囊網路的識別準確率，卷積層的數目是否為關鍵因素，是本論文研究的主題。在本論文中進行三項實驗：（1）一層到多層卷積層架構的膠囊網路，參數量都相近，和卷積神經網路（2層卷積層及2層池化層）參數量較膠囊網路多，是否有池化層的效果比較好，做驗證準確率的變化及分析，也比較了測試準確率的比較及分析；（2）在多個卷積層架構的膠囊網路和卷積神經網路（2層卷積層及2層池化層）比較不同卷積尺寸（3×3）、（5×5）、（7×7）、（9×9）做測試準確率比較和分析；（3）在膠囊網路中視覺化特定架構的特徵圖，且微調數字膠囊層中16個維度，重建出相似於原圖的圖片，代表膠囊網路有學到維度裡面所有的意義，透過視覺化能夠更清楚了解膠囊網路裡頭的奧秘。

關鍵字

深度學習；卷積神經網路；膠囊神經網路；動態路由；擠壓函數

並列摘要

In the past, convolutional neural networks have always had excellent performances to the recognition tasks in the computer vision field. Convolutional neural networks are composed of convolutional layers, pooling layers, and fully connected layers. In particular, convolutional layers are important in convolutional neural networks. It is mainly to capture different features for subsequent classification. The capsule network proposed by Hinton scholars in 2017 is a neural network model based on a new vector concept and dynamic routing algorithm. The architecture in original paper uses a two-layer convolutional layer. To improve the accuracy of capsule network recognition, whether the number of convolutional layers is the key is an experiment in this thesis.Three experiments are carried out in this thesis: (1) Capsule networks with one-to-multi-layer convolutional layer architectures have similar parameters. Convolutional neural networks (two-layer convolutional layer and two-layer pooling layer) have more parameters than capsule networks. Is there a pooling layer? The effect of the verification is better. The change and analysis of the verification accuracy rate are also compared and the comparison and analysis of the test accuracy rate; (2) Capsule network and convolutional neural network (2-layer convolutional layer and 2-layer pooling layer convolutional neural network) compare different convolution sizes (3×3), (5×5), (7×7), (9×9) for test accuracy comparison and analysis; (3) Capsule network Visualize the feature map of a specific structure, and fine-tune the 16 dimensions of the digital capsule layer to reconstruct a picture similar to the original image, which means that the capsule network has learned all the meanings in the dimensions, through visualization, the mystery of the capsule network can be understood more clearly.