A Multi-hierarchical Attention Connected Encoder-decoder Framework for Cluttered Object Segmentation and Thickness Prediction

In the steelmaking industry, adopting recycled steel instead of iron ore as raw material can significantly reduce enterprise costs. Accurate identification of the edge thickness of recycled scrap steel is crucial, as the purchase price of a batch of scrap steel depends on its edge thickness classification result. Edge segmentation of scrap steel images facilitates the calculation of edge thickness; however, the daily influx of large quantities of scrap steel deliveries urgently requires an intelligent and precise method for classifying cluttered scrap objects and establishing accurate pricing. While existing UNet-based neural networks, with their high accuracy and low computational complexity, are widely used in image segmentation, especially in resource-constrained scenarios, but they still struggle with cluttered scrap steel images. Key limitations include insufficient fine edge feature extraction, poor adaptability to uneven illumination and irregular scrap shapes, and ineffective integration of global and local feature information. To address these challenges, this paper proposes a novel U-shaped neural network model specifically designed for edge segmentation and thickness recognition of scrap steel images. The model employs a Swin Transformer encoder-decoder as its backbone and introduces a multi-head channel-wise attention connection mechanism for the UNet to enhance information flow between encoding and decoding stages. Multi-head attention enables parallel learning of semantic relationships, improving scrap steel region localization under complex backgrounds. Channel-wise attention dynamically adjusts feature channel importance, amplifying responses to thickness-related cues and enabling clearer edge segmentation in cluttered stacking scenarios. With a dual enhanced feature fusion architecture, the model effectively integrates shallow texture, mid-level shape, and high-level semantic features, improving global feature modeling while maintaining low computational complexity. The experimental results on scrap steel image dataset show that the proposed method outperforms conventional models, achieving a 1.21% to 12.86% higher IoU in edge segmentation and reducing the thickness prediction error by 1.87% to 24.52%, confirming its effectiveness and practical value.