OBJECT DETECTION AND IDENTIFICATION IN REAL TIME USING DEEP LEARNING

Abstract

Deep learning has majorly taken over the field of computer vision, especially with regard to spotting objects in images. Models based on Convolutional Neural Net- works, or CNNs, have totally altered the situation. Instead of spending hours and hours selecting the features manually, models just do it by themselves. Object detection is not only limited to recognizing objects in an image, It is also about determining the exact locations of them. This has become important for applications such as autonomous vehicles, medical imaging tasks, security cameras, and even smart shopping systems. However, there exists a problem. Deep learning models require massive, varied datasets in order to perform well. The process of collecting images and labeling. It is costly and time consuming. The place where data augmentation comes in. People use many methods such as image flipping, image rotating, cropping or brightness changing to extend their raw datasets. It is quite helpful, but those changes are very simple. They do not handle for example the case in which one object is in front of the other or that there are complex backgrounds or poor lighting or even that some categories are barely present in the data. Therefore, models are still struggling with the disorder that exists outside the lab. The idea was taken to create a more intelligent data augmentation pipeline. The idea is to make training data look more like authentic. It means that changes that matter most for the task should be added, context should be mixed in, and occlusions should be simulated. So, the model will learn better patterns, not get stuck on the training data and just be able to perform better in the real world.