To address the ill-posed inverse problem, reconstruction artifacts, and blurred boundaries in electromagnetic tomography (EMT) imaging of fine cracks on planar metal components, an image reconstruction network named SE_ASPP_UNet is proposed. The network integrates squeeze-and-excitation (SE) and atrous spatial pyramid pooling (ASPP) modules into the U-shaped network (U_Net) framework. The SE module enhances crack-related channel responses, while the ASPP module extracts multi-scale contextual features in the bottleneck layer, improving the representation of cracks with different lengths and orientations. A combined binary cross-entropy and Tversky loss is adopted to improve the reconstruction of slender crack structures. Based on a regular strip-like crack simulation dataset jointly constructed using COMSOL and MATLAB, comparative experiments with various reconstruction methods, loss-function ablation experiments, and network-module ablation experiments were conducted. SE_ASPP_UNet achieved an average Dice coefficient of 0.836 0 and an average connected component error (CC_Err) of 0.055 2 on the simulation test set, showing favorable region overlap and topological connectivity. After fine-tuning with a small number of measured samples, measured validation on 6061 aluminum alloy plates showed that the method could recover crack boundaries and spatial locations. Out-of-distribution experiments further indicated that the model retained a certain ability to recover the macroscopic propagation trend of irregular strip-like cracks unseen during training. These results suggest that SE_ASPP_UNet can support EMT image reconstruction of fine cracks in planar metals.