Genetic Algorithms for Optimising Context-based Neural Networks for Image Segmentation

Image segmentation is one of the major challenges in real-world computer vision applications. Context-embedded network models proposed for image segmentation have outperformed context-free models. However, optimized values of many parameters need to consider for such a complex network. The manual parameter selection process is ineffective and produces suboptimal performance for such a model. Therefore, we propose a context-based genetically optimized network model for image segmentation in this paper. Genetic algorithms enhance the performance of the deep network model by determining the best parameter values. The proposed three-level deep network is adaptable to image context by extracting visual and context-rich features and optimally integrating them to obtain final pixel labels for scene images. The genetic algorithm ensures optimal parameter values in all three levels to obtain a globally optimized network model to achieve the best segmentation results.

I. Introduction

One of the most basic yet difficult computer vision tasks is scene segmentation, which aims to separate and parse a scene image into various image regions related to semantic categories, including background texture (such as the mountain and the sky) and distinct objects (e.g., car, person, and bike) at the pixel level. The outcome of semantic segmentation maintains the very same matching resolution as the input image, calling it a dense prediction. Understanding images is an important step in vision systems that has countless uses. Finding patterns of interaction among class categories for segmenting an image into semantically meaningful categories is a major challenge for an image segmentation model. Accurate pixel categorization into object categories in complicated natural situations is a difficult task due to the visual heterogeneity across structured and unstructured items [l].