Transfer Learning Based Neural Network for Object Detection

Abstract

Computer vision is a field of study that focuses on how computers can interpret and analyze visual information from the world around us. This includes tasks such as image and video recognition, object detection, facialrecognition and scene reconstruction.Therefore, video analysis and understandingthe images has become necessary and challenging issue. This study aimed at presenting the model based on benchmark MSCOCO datasetof image data. Highly accurate object detection-algorithms and architectures such asFaster R-CNN, Mask R-CNN with backbone architectures Resnet, Inception and ResNeXt are fast yet highly accurate ones like YOLOv3 and YOLOv4 captures both low-level and high-level features. These models exhibit different behaviors in terms of network architecture, training methods, and optimization techniques, etc. Each and every object in an image is identified by the area object in a highlighted rectangular boxes and tag is assigned to each and every object.The accuracy in detecting theobjects is checked by different parameters such as accuracy, frames per second(FPS) and mean average precision (mAP). Also, the performance of the presented model with YOLOv4 andDarknet-53as a backbone architecture for transfer learningor fine-tuned for specific computer vision taskachieves 49.2% mAP, which outperforms the baseline by 3.8%.