Developing a system capable of recognizing objects and individuals in images through the application of convolutional neural networks

Abstract

The aim of this study was to design, test, and fully analyze a combined real-time system using convolutional neural networks (CNNs) with the ability to achieve general object detection and single (person) re-identification in unconstrained images simultaneously. The system proposed follows a hybrid two-stream design: the former stream uses an improved version of YOLOv5 to quickly and precisely detect multi-objects, and the latter stream is based on the modified ResNet-50 backbone that was trained using the triplet and ArcFace losses to learn highly discriminative identity representations. Other improvements are Squeeze-andExcitation attention blocks, widespread data augmentation, ImageNet training transfer, mixed-precision training and distributed, multi-GPU optimization. Strict testing using massive benchmark tasks (COCO, CelebA, cross-dataset), showed an average Average Precision (mAP 0.5:0.95) of 0.62 on object detection and a top-1 identification rate of 92 percent, which is a 817 percent improvement over baseline models. It was shown to high-level adversarial resistance, variations in illumination, and partial occlusions with real-time inference time of less than 200 ms on consumer-grade GPUs. Extensive studies regarding ablation and demographic equity also confirmed the role of every component and limited bias. Such findings make the suggested framework a very practical and deployable tool in a large number of applications, such as intelligent surveillance, assistive robotics, humanrobot interaction, forensic analysis, and smart environments. This work in the end provides a resultant reproducible, scalable and state-of-the-art pipeline that can be used as a strong building block to unified visual recognition systems in the next generation.