Indoor Scene Recognition using ResNet-18

摘要

Deep learning allows computational models consisting of multiple neural layers of processing to learn data representation at multiple abstraction levels. Such approaches have greatly strengthened state-of-the-art speech recognition, visual object recognition, scene recognition, NLPs, object detection, and many other fields such as drug discovery, distant surgeries and genomics. Scene Recognition is an area of visual recognition where we design and automate our system to recognize and identify the scene of the image. Automatic Scene Recognition or Scene Analysis is one of the hot topics in Deep Learning. If we look at the contributions of Deep Learning in this decade, we had come to know that Scene Recognition has been an obvious concern for scientists as it has great significance in security and surveillance too. Object recognition and recognition of indoor scenes plays a significant role in the cognition of service robots in the field. The development of deep learning has made fine-tuning of CNN (Convolutional Neural Network) on target datasets a common way of solving classification problems. Nonetheless, this approach cannot achieve adequate results easily for indoor scene classification due to over fitting when the datasets for scene preparation are inadequate. In order to compare techniques that participate to better accuracies, we have applied two different techniques to achieve our results i.e. Fine Tuning and concept of Freezing Layers. Within this project, a system of classification of the indoor scene is proposed to solve this issue. We are using ResNet-18 which contains 18 deep layers for classification. Furthermore, we are using transfer learning and performing classification on scenes based images.