AN IMPROVED MULTI-OBJECT INSTANCE SEGMENTATION BASED ON DEEP LEARNING

Abstract

Deep Learning (DL) networks have attracted growing interest and attention by researchers and scholars alike due to the growing importance of detecting and instance segmentation of objects in an image. Instance segmentation is a critical issue that requires further improvement due to the difficulties in adapting object detection and instance segmentation approaches. This paper presents an approach that overcome these issues by proposing a new approach based on the recent DL approach in addition to developing an approach for multi-object instance segmentation. The improved multi-object segmentation approach presented in this paper consists of three stages. Firstly, it improves the RestNet-101 (Residual Neural Network) backbone by connecting it to the convolution layer for each ResNet block. Secondly, the localization of multiple objects is improved by enhancing the Region Proposal Network (RPN), and thirdly, a complex instance segmentation approach is utilized. The result of this study based on a standard dataset, called the Common Object in Context (COCO) dataset (Lin et al., 2014), reveals that the suggested approach compared to other well-known segmentation approaches, has improved the instance segmentation process in terms of precision and training time.