A Multi-Scale Conv GCN-FCN Approach with Hybrid Optimization for Enhancing Object Detection in Occlusions

Abstract

Object detection is a fundamental problem in computer vision that involves identifying and localizing instances of objects within digital images. Real-time object detection systems often operate in dynamic environments where the scene are rapidly changing. Noise, poor lighting, background clutter and occlusion are some significant factors that impact the quality of object detection. In real-time object detection, objects can often be overlapped with each other. Occlusion is challenge in which an object is partially obscured by other objects, poses significant challenges for accurately detecting and localizing objects in videos. This common issue hampers the performance of object detection algorithms. Therefore, existing approaches to object detection often struggle with accurately localizing and detecting partially occluded objects. This research offers an in-depth analysis of the most recent state-of-the-art approaches for addressing the occlusion problem in object detection. Unlike previous occlusion detection methods, we introduce a novel approach that utilizes multi-scale context information to handle occlusion in object detection better. We proposed conv GCN-FCN with a hybrid optimization approach and evaluated on two standard datasets: MS COCO Dataset, and PASCAL VOC Dataset. We investigate the efficiency of our novel approach, by using a Graph convolution network (GCN) with a fully convolution network (FCN) and hybridization of the booster algorithm. Our proposed refined selfish herd optimizer enhance the neural network's performance that utilizes multi-scale context information to handle occlusion in object detection better. The experimental results demonstrate our proposed method significantly improves occlusion handling and also improves evaluation parameter such as accuracy, precision, F1 score and sensitivity etc. compared to previous CNN algorithms and methods.