我們的研究目標是開發促進自動視覺識別的方法。為了預測與影象相關的唯一或多重標籤，我們研究了用於監督特徵學習的不同型別的深度神經網路結構和方法。我們首先回顧了卷積神經網路的最新進展，旨在瞭解這一系列統計模型背後的歷史、現代結構的侷限性以及當前用於訓練深層CNN的新技術。我們工作的獨創性在於我們專注於低資料量任務處理的方法。我們採用不同的模型和技術在幾種資料集上獲得最佳精度，例如用於構建網頁API的中等食物食譜資料集（100k影象）或用於DSG線上挑戰的小規模衛星圖片（6000張）。我們還擬定了弱監督學習的最新技術，引入了能夠定位感興趣區域的不同型別的CNN。我們的最後一個貢獻是建立在Torch7之上的框架，用於在任何視覺識別任務和任何規模的資料集上訓練和測試深度模型。

The goal of our research is to develop methods advancing automaticvisual recognition. In order to predict the unique or multiple labelsassociated to an image, we study different kind of Deep Neural Networksarchitectures and methods for supervised features learning. We first draw up astate-of-the-art review of the Convolutional Neural Networks aiming tounderstand the history behind this family of statistical models, the limit ofmodern architectures and the novel techniques currently used to train deepCNNs. The originality of our work lies in our approach focusing on tasks with alow amount of data. We introduce different models and techniques to achieve thebest accuracy on several kind of datasets, such as a medium dataset of foodrecipes (100k images) for building a web API, or a small dataset of satelliteimages (6,000) for the DSG online challenge that we’ve won. We also draw up thestate-of-the-art in Weakly Supervised Learning, introducing different kind of CNNsable to localize regions of interest. Our last contribution is a framework,build on top of Torch7, for training and testing deep models on any visualrecognition tasks and on datasets of any scale.

1 卷積神經網路

2 深度CNN的遷移學習

3 弱監督學習

附錄A Overfeat

附錄B Vgg16

附錄C InceptionV3

下載英文原文地址：

更多精彩文章請關注微訊號：