Efficient Interactive Image Segmentation Using Multilayer Perceptron

Abstract

Interactive Image Segmentation (IIS) is important for applications that require precise segmentation with minimal user intervention, for example, in medical imaging, 3D object segmentation and video editing. However, the existing IIS methods are often hampered by ambiguity of interaction, leading to suboptimal performance of segmentation. We propose a novel IIS framework based on Multi-Layer Perceptron (MLP), which fuses clicks from users with feature representations extracted from the Transformer Encoder. We enhance the segmentation accuracy by using the adaptive click-based refinement and optimise with focal loss to address class imbalance. Extensive experiments on benchmark datasets including GrabCut, Berkeley, SBD, DAVIS and Pascal VOC show that our method outperforms state-of-the-art techniques in terms of segmentation accuracy and robustness. The proposed method achieves lower NoC@85 and NoC@90 values across several benchmark datasets, with particularly strong performance on GrabCut and BDD100K. We use a Vision Transformer-based architecture with more advanced training methods to improve feature extraction and boundary delineation. This guarantees reliable convergence of the model and good generalisation. Qualitative and quantitative evaluations validate that the framework is effective in real-world segmentation tasks. Our method closes the gap between interactive user instructions and deep learning and provides a practical and flexible solution for accurate segmentation in many applications. The proposed framework differs from existing transformer-based IIS methods by incorporating a Vision Transformer encoder and a lightweight MLP segmentation head and by using normalised focal-loss optimisation to enhance the interaction efficiency while maintaining high segmentation accuracy. The source code of the proposed approach is available at https://github.com/abhigoku10/IIS.git for reproducibility and further research.