CCCD: A Deep Deterministic Policy Gradient based Data Augmentation Technique using Change Detection and Automatic Augmentation for Counting a Crowd

Abstract

Abstract The necessity to address a difficult, significant, real-world image-based counting problem that cannot be adequately addressed with existing methodologies is what inspired this research. In order to overcome the difficulties mentioned above, we suggest a fresh method for teaching counting that builds on the earlier methods in several ways. A useful method for improving the performance of the counting model is data augmentation. The use of data augmentation can raise the amount and quality of training datasets, enhance model performance, and prevent the issue of data over-fitting. The phrase "data augmentation" refers to a collection of techniques used to increase the quantity and quality of training datasets so that Deep Learning models can be trained from them. Geometric transformations like color-space enhancement, kernel filters, combining pictures, and feature space augmentation are just a few of the image augmentation methods included in this examination. We propose a novel deep framework for counting, based on deep reinforcement learning. A pre-trained model with change detection is used before repeatedly attempting to build a Deep Deterministic Policy Gradient (DDPG)-based data augmentation strategy. After determining the optimum augmentation action for a given dataset, the augmented dataset is utilized to enhance the model. The outcomes of the experiment demonstrate that the automatic augmentation method may be utilized to produce adaptable augmentation strategies for counting models targeted at certain datasets.