Real-time instance segmentation algorithm based on mask activation and feature enhancement

With the widespread deployment of the Internet of Things, the demand for real-time environmental perception has become increasingly urgent. In this context, instance segmentation technology has emerged as a pixel-level scene perception method, garnering significant attention. This paper proposes a novel and efficient instance segmentation network designed for precise scene perception. In the decoding stage, we design a mask activation module to construct multi-layer weight matrices, with each layer directly activating a mask region of an instance, thereby achieving simplicity and efficiency. During the feature enhancement stage, we introduce two crucial modules to improve performance. Firstly, the global feature perception module models global dependencies through the self-attention mechanism, extending the network's receptive field. Secondly, the foreground feature capture module employs parallel convolutional kernels of various shapes and sizes to comprehensively explore foreground instance information. Experimental verification on the MS-COCO dataset demonstrates that our method achieves a better balance between accuracy and speed, and has potential in practical applications.