Image enhancement is an ongoing research problem that the community is addressing through the development of fusion algorithms. Such techniques typically involve the reconstruction of RGB images by removing environmental artifacts and enhancing desired features. Infrared imagery is also widely used to improve situational awareness in low visibility scenarios. Recently, learning-based approaches are used for fusion purposes to extract meaningful representations from images and capture latent features that could otherwise be inaccessible using conventional image processing algorithms. The inadequacies of RGB images in these algorithms’ pipelines are still obvious, despite the fact that the viability of RGB-Infrared image fusion has been thoroughly demonstrated in the literature. For example, RGB images often have artefacts like sudden changes in exposure or motion blur when the illumination changes or sudden changes in the scene. A novel imaging sensor operating in the visible light spectrum has been developed to address these issues. In this paper, we explore the cutting-edge paradigm of Neuromorphic Vision Sensors (NVS), a class of asynchronous analog imaging platforms that operate based on the change of pixel luminosity within a scene. When compared to frame-based counterparts, NVS enhances scene interpretation, processing time, reaction time, and power consumption. Deep-learning reconstruction networks are evaluated in this study to determine the applicability of existing state-of-the-art multi-modal image fusion techniques with the addition of NVS data rather than RGB data. As a benchmark, metrics such as signal-to-noise ratio (SNR) and pixel wise error are used.

Object detection is a critical task in computer vision, with applications ranging from robotics to surveillance. Traditional RGB-based methods often face challenges in low-light, high-speed, or high-dynamic-range scenarios, resulting in blurred or low-contrast images. In this paper, we present a novel algorithmic approach that fuses event data from event cameras with RGB images to improve object detection performance in real-time. Event cameras, unlike traditional frame-based cameras, provide high temporal resolution and dynamic range, capturing intensity changes asynchronously at the pixel level. Our method leverages the complementary strengths of event data and RGB images to reconstruct blurred images while retaining contrast information from the original data. We propose an algorithmic pipeline that first fuses event data and RGB images, followed by a reconstruction step to generate enhanced images suitable for object detection tasks. The pipeline does not rely on deep learning techniques, making it computationally efficient and well-suited for real-time applications. To validate the effectiveness of our approach, we compare its performance against the popular YOLO benchmarks for object detection tasks. Moreover, we assess real-time metrics to demonstrate the practicality of our method in time-sensitive applications.

No abstract