The Tooploox team will be participating in the ACIIDS conference, submitting three research papers that contribute to science by introducing Schrödinger bridges, supporting satellite imagery, and helping machines build better length estimations.
The Asian Conference on Intelligent Information and Database Systems has been held since 2009 in various locations in Asia and the Middle East. The event has taken place in Bali, Indonesia; Kuala Lumpur, Malaysia; Hue City, Vietnam; and Ras Al Khaimah, the United Arab Emirates; among others.
The conference focuses on scientific research around Databases, Information Systems, Artificial Intelligence, and everything in between. The event is also an opportunity for researchers from all around the world to exchange ideas, knowledge, and meet each other to let knowledge flow.
This year the conference will see three Tooploox-affiliated papers published.
Image enhancement with Boosted Schrödinger Bridge
The first paper focuses on using novel AI-based techniques to improve image quality.
Schrödinger bridges are a novel class of diffusion-based methods adapted to image restoration tasks, such as super-resolution, denoising, and image colorization. With this tool, old, low-quality or low-resolution images can be granted new life (or become accessible to new generations of viewers) with new found quality delivered by machine learning algorithms.
The downside of existing methods
Unlike traditional denoising diffusion models, Schrödinger bridges are designed so that their stationary distribution is centered on the input image to be enhanced rather than pure noise. However, their optimization process still relies on mean squared error (MSE) criterion between the prediction and the applied noise.
In this work, the researchers propose an improvement to Schrödinger bridge models by enriching their loss function with additional information. Specifically, we introduce dynamically weighted style-based loss that leverages a pre- trained model to extract high-level feature maps for comparison against ground-truth embeddings. The enhancement improves the performance of Schrödinger Bridge architectures on a range of benchmarks, demonstrating the effectiveness of our approach in advancing image enhancement capabilities.
Practical applications
The tool can be used to improve the quality of images of all sorts. For example, the system may restore the quality of old photographs to make one’s ancestors recognizable again. This also applies to upscaling movies, so new audiences can enjoy cult classics and masterpieces of cinema that have lost their allure due to blurred images and low resolution.
Satellite Based One-Class Classifier Models with Augmentation for Efficient Mineral Exploration
The second paper prepared for the ACIIDS conference is about leveraging the power of AI algorithms in mineral detection using hyperspectral data gathered by satellites.
What is hyperspectral data
Traditional images represent only the data available in the visible light spectrum – those detectable by the human eye. But compared to machine sensors, our biological apparatus is flawed. That’s why gathering information using satellites equipped with sensors which go beyond our own senses is so effective and gives users way more information than by using the eyes alone.
This study proposes a novel approach to the classification of hyperspectral data using a one-class classifier, focusing on a single target class (mineral spectra) to address challenges such as limited training data and environmental interference – for example the atmosphere or vegetation.
Tackling the limited dataset
To mitigate the issue of sparse labeled samples, we employed data augmentation techniques to synthetically expand the training set. We compared the performance of augmented and non-augmented datasets, optimizing for F-Score to balance false positives and false negatives.
The experimental results show that both Recall and F-Score were significantly higher for the augmented datasets. Specifically, the augmented datasets achieved higher Recall, indicating an improved ability to identify potential mineral regions, and a higher F-Score, demonstrating a better balance between precision and recall compared to non-augmented datasets.
These results, validated in the Cuprite region, confirm the effectiveness of data augmentation in improving mineral area detection, enhancing the reliability of hyperspectral imaging for mineral exploration despite challenges such as limited labeled data and environmental factors.
Practical applications
This technique can be used to detect interesting mineral resources using satellite imagery. The tool can be used to narrow areas of interest and save significant resources by avoiding areas with little to no minerals of interest.
Direct detection of elongated object geometry via a centerline-based representation
This paper examines the AI’s ability to detect not only an object itself, but also its geometry – whether the object is elongated or not.
Object detection aims to locate objects in images by classifying them and assigning them bounding boxes. However, this representation is insufficient when details of the object’s geometry are of interest.
In such cases, more computationally intense methods such as instance segmentation are necessary. Furthermore, bounding boxes do not distinguish between heavily occluding objects.
In this work, we propose a more geometrically descriptive representation: a centerline. We’ve trained an object detection model to predict centerlines instead of bounding boxes. This allows us to obtain some geometric properties of the objects directly in detection, without the need for segmentation. We also find the proposed approach to improve overall detection performance over the bounding-box baseline.
Practical applications
In many cases, the practical application of computer vision techniques requires some degree of measurement or estimation of the spatial relationship between depicted objects. Also it can be used to automatically sort elongated objects present in images – for example, for automating agricultural processes.
Summary
As is shown in the papers above, the conference is a great opportunity to exchange knowledge and experience regarding multiple aspects of computer vision applications. It is a place where experts on satellite imagery, microscopy, and everything in between may meet and discuss their challenges and opportunities.
The conference will take place on 23-25 April 2025, Kitakyushu, Japan. More detailscan be found on the event website.
