ForAug: Mitigating biases in classification with controlled compositions

In the field of image classification, artificial intelligence models often leverage compositional biases —such as the centered position of objects, their characteristic scale, or the presence of specific backgrounds— to achieve high accuracy in controlled environments. However, these same correlations make them fragile when faced with distribution shifts, a critical problem in real-world applications. To address this, ForAug emerges, a data augmentation scheme based on controlled compositions that decomposes each image into foreground and background, recombining them to break spurious correlations. This technique allows explicit manipulation of the object's position, scale, and background, reducing unwanted shortcuts and improving model robustness. From a business perspective, integrating strategies like ForAug into computer vision systems can make the difference between a model that only works in a demo and one that performs in production. At Q2BSTUDIO we offer artificial intelligence for businesses, helping to implement robust classification solutions through custom software and custom applications tailored to each use case. We combine AWS and Azure cloud services to scale image processing and business intelligence services with Power BI to visualize model performance. Additionally, our AI agents can automate bias detection in datasets, while cybersecurity ensures the integrity of sensitive data. The key lies not just in training models, but in designing augmentation pipelines that generate real generalization. ForAug exemplifies how fine-grained control over training data can mitigate positional and background biases, improving by up to 19 percentage points on distribution shift benchmarks. At Q2BSTUDIO we apply these principles in computer vision projects, integrating advanced composition and augmentation techniques so that models do not memorize spurious correlations, but learn invariant features. This way, companies can deploy more reliable systems, from automated visual inspection to medical image analysis, backed by a team of experts in software development and cloud technologies.