Deep Learning Techniques for Scene Recognition

在过去的几年里，深度学习无疑改变了机器学习和计算机视觉的面貌，但在某些方面，研究领域已经采取了倒退的一步，专注于更简单的（有点人为的）图像分类问题，而不是整个“场景理解”。本博士的基础研究将着眼于在整个语义场景理解中重新点燃一些旧的想法，并根据使用深度学习取得的进展，用新的想法丰富这些想法。人类有能力快速分析和理解图像或视频捕捉中所呈现的情况背后的含义。这种能力是必不可少的，因为它使我们能够在日常生活中执行不同的任务。我们能够识别物体、特征和纹理，对它们进行正确分类，然后为整个场景赋予意义。尽管机器在物体识别任务上取得了很大的进步，但要想在场景识别任务中达到与人类相当的效率和准确性，它们还有很长的路要走。与物体不同，场景是由整个图像来表示的，而不仅仅是其中的某些部分。如今，高阶语义场景识别是视觉计算领域最有趣的话题之一，我们认为训练机器识别和讲述图像背后的故事可以极大地造福我们的生活。这一研究课题的重要性可以在监控、驾驶辅助和人机交互等领域观察到。

Deep learning has undoubtedly changed the face of machine learning & computer vision over the last few years, but in some ways, the research field has taken a retrograde step by focusing on simpler (and somewhat contrived) problems of image classification rather than of whole 'scene understanding'. The fundamental research of this PhD will look to reignite some of the older ideas in whole semantic scene understanding, and enrich these with new ideas, in the light of the advances made using deep learning.Humans have the ability to quickly analyse and understand the meaning behind a situation deputed in an image or video-capture. This capacity is essential because it allows us to carry out different tasks in our every-day lives. We are able to recognize objects, features and textures, classify them correctly and then assign meaning to the whole scene. Although machines have made great progress in object recognition tasks, they still have a long way to go until they will be able to perform with human comparable efficiency and accuracy in scene recognition tasks. Unlike objects, scenes are represented by entire images, not just certain parts of it. Nowadays, high order semantic scene recognition is one of the most interesting topics in the visual computing field and we consider that training machines to recognize and tell the story behind an image can greatly benefit our lives. The importance of this research topic can be observed in areas such as surveillance, driving assistance and human-machine interaction.