Interpreting Convolutional Neural Networks (CNNs) is critical for safety-sensitive applications such as healthcare and autonomous systems. Popular visual explanation methods like Grad-CAM use a single convolutional layer, potentially missing multi-scale cues and producing unstable saliency maps. We introduce Winsor-CAM, a single-pass gradient-based method that aggregates Grad-CAM maps from all convolutional layers and applies percentile-based Winsorization to attenuate outlier contributions. A user-controllable percentile parameter $p$p enables semantic-level tuning from low-level textures to high-level object patterns. We evaluate Winsor-CAM on six CNN architectures using PASCAL VOC 2012 and PolypGen, comparing localization (IoU, center-of-mass distance) and fidelity (insertion/deletion AUC) against seven baselines including Grad-CAM, Grad-CAM++, LayerCAM, ScoreCAM, AblationCAM, ShapleyCAM, and FullGrad. On DenseNet121 with a subset of Pascal VOC 2012, Winsor-CAM achieves 46.8% IoU and 0.059 CoM distance versus 39.0% and 0.074 for Grad-CAM, with improved insertion AUC (0.656vs. 0.623) and deletion AUC (0.197vs. 0.242). Notably, even the worst-performing fixed $p$p-value configuration outperforms FullGrad across all metrics. An ablation study confirms that incorporating earlier layers improves localization. Similar evaluation on PolypGen polyp segmentation further validates Winsor-CAM's effectiveness in medical imaging contexts. Winsor-CAM provides an efficient, robust, and human-tunable explanation tool for expert-in-the-loop analysis.

Abstract Historical census records convey information that is key to perform genealogical research and demographic studies. Given the large number of documents of this type that exist, it is crucial to research methods that allow the automatic extraction of information from this type of document. In this work, we present a new corpus of this kind, comprising 535 historical census tables from French archives. Alongside this dataset, we have assessed three different baseline methods for information extraction. The first two methods employ a traditional sequential approach, where table rows are detected before extracting information. The third baseline uses an end-to-end model that directly extracts information from the table images without prior row detection. Our results demonstrate the effectiveness of all three baselines in tackling the information extraction task.