• Propose a novel noise-reduction-oriented super-resolution framework tailored for precision agriculture applications. • Introduce a Gaussian-based saliency map generation method to guide object-aware SR reconstruction. Design a multitask learning architecture that jointly predicts high-resolution images and saliency maps to enhance edge and contour understanding. • Integrate a saliency-guided segmentation post-processing technique to automatically select optimally enhanced high-resolution outputs. • Achieve state-of-the-art PSNR performance in noise reduction across four benchmark datasets and one agriculture-related dataset. Low-resolution images significantly degrade the performance of vision systems in real-world settings. While super-resolution techniques improve image details, they often introduce additional noise, complicating tasks like detection and recognition. Thus, it is crucial to develop methods that enhance image quality while reducing noise. This paper presents a novel solution focused on noise reduction, integrating Super-Resolution (SR) reconstruction, Multitask Learning (MTL), and Saliency-Guided Segmentation (SGS) for effective post-processing. When generating high-resolution images, existing SR methods struggle with noise at object boundaries and edges. Our model learns a supplementary task to create blending saliency objects into high-resolution images, thus improving the understanding of object boundaries, edges, and contours, while leading to notable noise reduction. Experimental results show that SUNRISE achieves the highest PSNR on benchmark datasets, including SET5 (33.44 dB), SET14 (32.17 dB), BSD100 (31.87 dB), and Urban100 (31.81 dB), outperforming state-of-the-art methods such as FxSR, DualFormer, SROOE, and WGSR by up to 1.5 dB. On the Saffron Flower Dataset, SUNRISE achieves a PSNR of 33.36, an SSIM of 0.824, MSE of 35.957, and LPIPS of 0.232, demonstrating superior pixel-level fidelity, structural preservation, and perceptual similarity in agricultural images. These results confirm that SUNRISE effectively reduces noise while preserving fine details, thereby enhancing downstream vision tasks like object detection. The code is available at https://github.com/tmtgssi/SUNRISE

Abstract The capability to integrate Internet of Things (IoT) technologies into business processes (BPs) has emerged as a transformative paradigm, offering unprecedented opportunities for organisations to enhance their operational efficiency and productivity. Interacting with the physical world and leveraging real-world data to make more informed business decisions is of greatest interest, and the idea of IoT-enhanced BPs promises to automate and improve business activities and permit them to adapt to the physical environment of execution. Nonetheless, combining these two domains is challenging, and it requires new modelling methods that do not increase notation complexity and provide independent execution between the process and the underlying device technology. In this work, we propose FloBP, a model-driven engineering approach separating concerns between the IoT and BPs, providing a structured and systematic approach to modelling and executing IoT-enhanced BPs. Applying the separation of concerns through an interdisciplinary team is needed to ensure that the approach covers all necessary process aspects, including technological and modelling ones. The FloBP approach is based on modelling tools and a microservices architecture to deploy BPMN models, and it facilitates integration with the physical world, providing flexibility to support multiple IoT device technologies and their evolution. A smart canteen scenario describes and evaluates the approach’s feasibility and its possible adoption by various stakeholders. The performed evaluation concludes that the application of FloBP facilitates the modelling and development of IoT-enhanced BPs by sharing and reusing knowledge among IoT and BP experts.