MilInst: Enhanced Instance Segmentation Framework for Military Camouflaged Targets Using Sparse Instance Activation

Abstract

In this study, an improved end-to-end framework for instance segmentation of military camouflaged targets, referred to as MilInst, is proposed. The framework builds upon SparseInst method developed by Cheng et al. (2022). Several improvements are introduced to enhance the model’s performance. First, Receptive Field Enhancement Module (RFEM) is employed to capture broader contextual information. Additionally, Feature Merging Module (FMM) is utilized to eliminate feature noise through the implementation of a matrix decomposition method. Furthermore, a novel linear dynamic bipartite matching approach is proposed, facilitating a smooth transition from one-to-one matching to one-to-many, and eventually achieving accurate to one-to-one matching. Experimental results demonstrate the effectiveness of MilInst algorithm. Comparisons with a selected real-time instance segmentation baseline model reveal superior performance, with MilInst framework achieving the highest mean Average Precision (mAP) index of 84.8% on a self-made dataset.

Keywords

• Instance segmentation
• deep learning
• military camouflaged targets