Presentations

• “Does Baxter Dream of Electric Beans?” (oral) – AgriFoRwArdS CDT Summer School 2023 [March 2023] – Lincoln, UK.
• “Training strategies for domain generalisation in object detection for autonomous driving” (poster) – Towards Autonomous Robotic Systems (TAROS) 2023 / AgriFoRwArdS CDT Annual Conference 2023 / Joint Robotics CDT Annual Conference 2023 [September 2023] – Cambridge, UK.
• “Understanding and mitigating the problem of highlights in remote sensing with application to coastal surveying​” (oral) – AgriFoRwArdS CDT Quarterly PhD Research Progress Meeting [January 2024] – Lincoln, UK.
• “Unifying Path and Center-Surround Retinex Algorithms” (oral) – London Imaging Meeting (LIM) 2024: Image capture, from photons to photos [June 2024] – London, UK.
• “Jacobi Center-Surround Retinex and Gauss-seidel Retinex” (poster) – AgriFoRwArdS CDT Quarterly PhD Research Progress Meeting [June 2024] – Cambridge, UK.
• “Jacobi Center-Surround Retinex and Gauss-seidel Retinex” (poster) – AgriFoRwArdS CDT Annual Conference 2024: Robots in Action [July 2024] – Norwich, UK.
• “Domain Generalisation for plant/weed detection” (oral) – AgriFoRwArdS CDT Summer School: Robotic Phenotyping [July 2024] – Wageningen, The Netherlands.
• “Jacobi Center-Surround Retinex and Gauss-seidel Retinex” (poster) – AgriFoRwArdS CDT Summer School: Robotic Phenotyping [July 2024] – Wageningen, The Netherlands.
• ‘Jacobi Center-Surround Retinex and Gauss-seidel Retinex’ (poster) – Centre for Environment, Fisheries and Aquaculture Science (Cefas) PhD Student Conference 2024 [December 2024] – Lowestoft, UK.

About me

I hold a bachelor’s and a master’s degree in Biosystems Engineering. Prior to joining the CDT, I worked as a researcher in the R&D division at Packman Company. I also have eight years of experience as a mechanical and agricultural engineer across various companies.

Throughout my career, I have developed strong expertise in several areas, including programming, solid mechanical design, simulation and 3D modelling, structural and mechanism analysis, agricultural machinery design, agricultural mechanization technologies, and manufacturing processes.

I am joining the CDT to pursue my interest in computer vision and agri-robotics, and to explore how these technologies can advance modern agricultural systems.

MSc Project

Training strategies for domain generalisation in object detection for autonomous driving

This project aims to identify hierarchical training strategies that improve the ability of object detection models in autonomous driving to generalise across different domains. Hierarchical learning as a training strategy is used for domain generalisation purposes. It applies to the training step of the model and helps the algorithm find domain invariant features for object detection tasks. This project utilises the Faster R-CNN as the detection algorithm, ACDC (adverse conditions dataset city) as training data, and the Cityscape dataset as the test dataset.

PhD Project

Understanding and mitigating the problem of highlights in remote sensing with application to coastal surveying

In today’s increasingly digitized world, computer vision plays a vital role across a wide range of sectors, from agriculture to coastal monitoring. However, these systems often face challenges when operating in environments affected by glare like sun reflection from the water surface. The information clipped in these areas leads to inaccuracies in environmental classification and object detection, limiting the effectiveness of computer vision applications in real-world scenarios.

In collaboration with the Centre for Environment, Fisheries and Aquaculture Science (CEFAS), our project seeks to address this critical challenge and enhance vision systems used for remotely piloted aircraft. We are developing a novel Retinex algorithm designed to recover lost information in clipped regions, improving the accuracy and reliability of computer vision systems operating in complex, high-glare environments.

The input to our Retinex algorithm is a high dynamic range (HDR) image, created by combining several short-exposure images. This approach captures a wider dynamic range and reduces the probability of clipped areas where information is lost.

The Retinex algorithm aims to decouple illumination from reflectance and eliminate illumination.  It mimics the behaviour of the human visual system, which can discount illumination and consistently perceive the true colour of objects under different lighting environments. This method can eliminate gradual changes in light and recover information in bright areas. We reframe the common McCann Retinex to a centre-surround structure, which closely aligns with the processing mechanisms of the human retina.

Compared to other Retinex methods, our approach offers several advantages: greater compatibility with the human visual system, faster convergence, higher-quality results, and fewer artefacts. Additionally, unlike deep learning approaches, our method does not require any training and performs effectively across a wide range of datasets.

Afsaneh’s PhD project is being carried out in collaboration with CEFAS, under the primary supervision of Prof Graham Finlayson.