상세정보

In this paper, we focus on learning multi label image classifier using only single label image dataset. Because a multi-label dataset requires annotation of all objects and concepts present in the image, it is time-consuming and effort-consuming to create a dataset, and most datasets that are readily available in real life are mostly single-label datasets. In this paper, we describe how to learn the Convolutional Neural Network (CNN) model using only single-labeled images, without creating or retrieving direct multiple label data.
We propose a Multiple Category Label Predictor (MCLP) to predict multiple labels from a single label image set in order to learn CNN as a multi-label prediction model with a single label image set alone. MCLP divides a set of single-label images into N categories, and then learns each category independently. The Category Label Predictor predicts the label of each category, and eventually performs the multi-label image classifier learning using the labels with high probability in all Category Label Predictors. 
We also propose a new loss function for learning of a multi-label image classifier using a single-label dataset as proposed. We conducted our experiments using our LG Image Dataset to evaluate the proposed method.

1. Introduction 1
2. Related Work 3
3. Method 4
3.1. Convolutional Neural Network 5
3.1.1. MobileNetV2 5
3.1.1.1. Depthwise Separable Convolution 6
3.1.1.2. Hyper Parameters : Width & Resolution Multiplier 8
3.2. Model Architecture 9
3.3. Inference 11
3.4. Loss Function 12
4. Experiments 14
4.1. Dataset 14
4.2. MCLP’s Hyper Parameters 15
4.3. Main CNN Model’s Hyper Parameters 15
4.4. Result 16
5. Conclusion 24