An improved YOLACT algorithm for instance segmentation of stacking parts

Abstract

Instance segmentation is a very important task for a variety of applications. Instance segmentation for stacking objects is a challenge for computer vision. To overcome the challenge, we propose an improved YOLACT (You Only Look At CoefficienTs) algorithm. To improve the accuracy of feature extraction, detection and segmentation in a densely stacking scene, a Multi-Level Feature Fusion and Channel Attention Mechanism Module (MLCA) are integrated with YOLACT's backbone. Further, to expand the receptive field without compromising image quality, we substitute the conventional Feature Pyramid Network (FPN) with an Attention-guided Context Feature Pyramid Module (AC-FPN). The effectiveness of the improved YOLACT algorithm is validated through extensive experiments on a customized dataset of stacking mechanical parts. Results demonstrate that the improved YOLACT algorithm significantly surpasses the other algorithms in detection and segmentation without notably increasing computing time.