Due to the limited depth of field (DoF) of optical microscopes, wear particles of varying thicknesses and sizes cannot be simultaneously presented in sharp focus within a single image, leading to potential misidentification of defocused particles in ferrograph analysis. To address this issue, an end-to-end unsupervised multi-focus ferrograph image fusion model, WearIF, is proposed, which takes a sequence of images as input and outputs an all-in-focus image. First, low-resolution focus weight maps are obtained using a bilinear downsampling operation and a multi-scale dense focus feature extraction network (MDFFEN). These maps are then refined through a convolutional guided filter network to generate high-resolution focus weight maps. Finally, the maps are weighted and summated with the source images to generate an all-in-focus ferrograph image. Moreover, a joint content and gradient-based unsupervised loss function is designed to train WearIF, with attention to image structure, texture details, and brightness balance. Experimental results show that WearIF retains more information from the source images and produces fusion results that are more natural and realistic compared to current deep learning-based fusion methods. The proposed model effectively reconstructs the morphology of defocused wear particles in ferrograph images, providing a solid foundation for ferrograph image analysis.

全文下载：End-to-End Deep Guided Learning for Reconstruction of All-in-focus Ferrograph Image.pdf