MFCP-Net: Multiscale Feature Correlation Perception Network for Image Segmentation Based on Background Prior

Due to their exceptional capacity to extract features, methods based on deep learning have demonstrated good performance in industrial vision inspection applications. However, the generalizability and robustness of inspection models are challenged by factors such as insufficient data, great differences in data types, and complex illumination. To address these problems, multiscale feature correlation perception network (MFCP-Net), a semantic segmentation model with variable topology, is proposed to segment images with prior background. MFCP-Net uses a Siamese structure with shared weights to extract the multiscale features of the template and test images. A feature correlation perception (FCP) block is designed to measure the correlation of latent representation between the template and test images at multiscales, using background information more effectively. In addition, MFCP-Net provides different topological structures for the training and testing phases. A background auto-correlation branch is included during the testing phase to eliminate false positives observed in the background. MFCP-Net was compared with the state-of-the-art models on the proposed automobile instrument detection system, demonstrating that it offers superior accuracy and generalizability, especially in the case of insufficient data and big variances across data.