Structured Filter Pruning applied to Mask R-CNN: A path to efficient image segmentation

This study investigates the optimization of two-stage object recognition systems, integral to a variety of applications, through advanced machine learning techniques. While such systems achieve high accuracy, they often suffer from over-parametrization and excessive computational demands. Addressing this, a pruning method tailored for complex architectures like Mask R-CNN was developed, aimed at reducing model complexity and storage without sacrificing accuracy. Employing a Global Kernel Level Filter Pruning strategy, identified by the L1-Norm, the approach strategically eliminated non-essential parameters post-training. This method maintains high recognition accuracy; the experiments demonstrated that accuracy remained robust up to 40\% pruning, achieving a compression ratioof 1.25 and an IoU@0.5 of 0.72. The findings offer valuable insights for the field, presenting a nuanced approach to enhancing Neural Network efficacy while maintaining performance, thereby contributing to the ongoing dialogue in AI optimization.