Neural Netw. 2025 Nov 14;196:108325. doi: 10.1016/j.neunet.2025.108325. Online ahead of print.
ABSTRACT
We conduct a comprehensive study on the quality inspection of lithium battery materials, which evaluates material conformity by analyzing particle sizes in scanning electron microscope (SEM) images. Currently, enterprises rely heavily on manual annotation to complete this task. However, manual annotation is labor-intensive and prone to subjective errors. To address these challenges, we reformulate the quality inspection task as the lithium battery materials SEM image segmentation (LBMS) task and aim to resolve it using artificial intelligence technology. To this end, we collect and construct a dedicated SEM image dataset for the LBMS dataset, called LBMS dataset. Then we propose a specialised model for the LBMS task, named LBMS-SAM. Specifically, we design an edge feature extraction module based on Sobel and Gabor convolutions (GSEFE), which aims to accurately extract and enhance image edge information. Additionally, We design a multi-layer denoised features fusion module (MDFF) that uses wavelet transform to denoise the output features of each global attention layer in the ViT model. The denoised features from different layers are then fused, enabling efficient extraction of global contextual information and suppressing noise introduced by the ViT architecture. The proposed model introduces minimal additional parameters, and extensive experiments on the LBMS dataset demonstrate that LBMS-SAM outperforms state-of-the-art (SOTA) methods across all relevant evaluation metrics.
PMID:41289643 | DOI:10.1016/j.neunet.2025.108325
