Segmentation of Instances in an Image with Custom Neural Networks

Image segmentation is a computer vision problem that involves splitting an image into particular regions, enabling precise object identification and analysis. While conventional U-Net architectures have excelled in acquiring hierarchical representations through the harmonious amalgamation of multi-scale features via symmetric contracting and expansive pathways, the precise delineation of object boundaries remains a formidable challenge. Within the context of this research paper, we introduce an attention-based custom U-Net model meticulously tailored to address the aforementioned challenges. This model achieves good performance metrics, including a maximum Intersection over Union (IoU) score of 75% and a minimal loss of 1.5 when operating at a downsampling rate of 64. These results mark a meaningful step forward in the field of image segmentation, providing insights into potential improvements in object boundary precision and contributing to ongoing progress in various application domains.