To address cross-well distribution shift, unstable boundary prediction, and missed detections in thin interbedded intervals during reservoir identification from multi-well logging data, this study proposes a DeepTabNet hybrid network method for fine-grained reservoir identification. Five well-log curves, namely gamma ray, deep resistivity, bulk density, neutron porosity, and sonic transit time, are selected as the input features, and vertical contextual information is introduced through robust normalization and sliding-window sample construction. DeepTabNet retains the step-wise sparse feature selection idea of the original TabNet, while further introducing feature embedding and multi-head self-attention to model high-order nonlinear inter-log interactions, together with a one-dimensional convolutional texture branch to extract local continuous morphological patterns along depth. Experiments are 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 PR-AUC of 0.905 in independent cross-well testing. It consistently outperforms the rule-based method, random forest, BiLSTM, TCN, and GNN baselines. Ablation experiments further confirm that multi-head self-attention, Sparsemax sparse gating, and the texture branch all contribute positively to performance improvement. The proposed method effectively improves reservoir identification accuracy, boundary stability, and cross-well generalization, while also providing good interpretability and practical engineering potential.