HMNet: High-Resolution Mamba Network for Semantic Segmentation of Surface Defects

Surface defect detection is widely used in various industrial scenarios. However, due to the characteristics of surface defects such as low contrast, blurry edges, and differences among individual defects, the detection task remains a significant challenge. Traditional semantic segmentation tasks based on Convolutional Neural Networks or Vision Transformers often underperform in establishing long-distance dependencies, while the latter suffers from high computational costs due to its complex second-order computations. Inspired by the Mamba state space model, we propose a high-resolution network (HMNet) based on the Mamba model. The main body of HMNet employs a special triangular structure, maintaining the size of the network output at 1/4 of the original image size, effectively focusing on the semantic features of fine-grained targets. We introduce the Contextual Mamba Block(CM Block) to construct the core sampling module within the encoder, aggregating positional information along both the horizontal and vertical directions based on the state space model, aiming to enhance the model's direction perception and global information fusion capabilities. We use experimental results from a private dataset of mobile phone glass screens and two publicly available datasets of surface defects on automotive parts and surface defects on hot rolled steel to demonstrate the effectiveness of HMNet for surface defect detection segmentation and its feasibility in industrial applications.