To address cross-well distribution shift, unstable boundary prediction, and missed detections in thin interbedded intervals during reservoir identification from multi-well logging data, a DeepTabNet hybrid network method for fine-grained reservoir identification was proposed. Five well-log curves, namely gamma ray, deep resistivity, bulk density, neutron porosity, and sonic transit time, were uniformly selected as the input features, and vertical contextual information was introduced through robust normalization and sliding-window sample construction. In terms of model architecture, DeepTabNet preserves the step-wise sparse feature selection strategy of the original TabNet while incorporating feature embedding and multi-head self-attention to capture high-order nonlinear interactions across logging curves. In addition, a one-dimensional convolutional texture branch was constructed to extract locally continuous morphological features along the depth direction, and the two types of features were ultimately fused to output reservoir probabilities. Experiments were conducted on logging data from 16 wells and evaluated using well-based splitting and five-fold cross-validation. The results show that DeepTabNet achieves an accuracy of 0.905, an F₁-score of 0.862, and a PR-AUC of 0.914 on the validation set, while obtaining an average accuracy of 0.899 and a PR-AUC of 0.905 in independent cross-well testing, consistently outperforming the rule-based method, random forest, BiLSTM, TCN, and GNN baselines. Ablation studies further demonstrate that multi-head self-attention, Sparsemax sparse gating, and the texture branch all make positive contributions to performance improvement. The results indicate that the proposed method can effectively enhance reservoir identification accuracy, boundary stability, and cross-well generalization, while also exhibiting good interpretability and practical engineering potential.