RETRACTED ARTICLE: A multi-filter deep transfer learning framework for image-based autism spectrum disorder detection

Autism Spectrum Disorder (ASD) affects approximately [Formula: see text] of the global population and is characterized by difficulties in social communication and repetitive or obsessive behaviors. Early detection of autism is crucial, as it allows therapeutic interventions to be initiated earlier, significantly increasing the effectiveness of treatments. However, diagnosing ASD remains a challenge, as it is traditionally carried out through methods that are often subjective and based on interviews and clinical observations. With the advancement of computer vision and pattern recognition techniques, new possibilities are emerging to automate and enhance the detection of characteristics associated with ASD, particularly in the analysis of facial features. In this context, image-based computational approaches must address challenges such as low data availability, variability in image acquisition conditions, and high-dimensional feature representations generated by deep learning models. This study proposes a novel framework that integrates data augmentation, multi-filtering routines, histogram equalization, and a two-stage dimensionality reduction process to enrich the representation in pre-trained and frozen deep learning neural network models applied to image pattern recognition. The framework design is guided by practical needs specific to ASD detection scenarios: data augmentation aims to compensate for limited dataset sizes; image enhancement routines improve robustness to noise and lighting variability while potentially highlighting facial traits associated with ASD; feature scaling standardizes representations prior to classification; and dimensionality reduction compresses high-dimensional deep features while preserving discriminative power. The use of frozen pre-trained networks allows for a lightweight, deterministic pipeline without the need for fine-tuning. Experiments are conducted using eight pre-trained models on a well-established benchmark facial dataset in the literature, comprising samples of autistic and non-autistic individuals. The results show that the proposed framework improves classification accuracy by up to [Formula: see text] points when compared to baseline models using pre-trained networks without any preprocessing strategies - as evidenced by the ResNet-50 architecture, which increased from [Formula: see text] to [Formula: see text]. Moreover, Transformer-based models, such as ViTSwin, reached up to [Formula: see text] accuracy, highlighting the robustness of the proposed approach. These improvements were observed consistently across different network architectures and datasets, under varying data augmentation, filtering, and dimensionality reduction configurations. A systematic ablation study further confirms the individual and collective benefits of each component in the pipeline, reinforcing the contribution of the integrated approach. These findings suggest that the framework is a promising tool for the automated detection of autism, offering an efficient improvement in traditional deep learning-based approaches to assist in early and more accurate diagnosis.