<p>This study presents a comprehensive systematic review of deep learning (DL) methods for wood image recognition and defect detection, synthesizing 168 peer-reviewed publications from 2015 to 2025 following the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) framework. The review addresses three research questions concerning the evolution of model architectures, the role of preprocessing and segmentation, and the technical limitations that hinder practical deployment. Convolutional Neural Networks (CNNs) remain dominant, accounting for 73% of studies due to their robustness and computational efficiency, while Transformer-based models, representing 12% of the literature, achieve accuracy up to 99.8% but require higher computational cost and larger datasets. Overall, DL models report classification accuracies between 95 and 99.8%, and defect-detection frameworks (such as YOLOv5/v8) reach mAP values above 95%, with several lightweight architectures achieving real-time inference on edge hardware. A key novelty of this review is the cross-modal comparison of RGB (Red, Green, Blue), NIR (Near-Infrared), SEM (Scanning Electron Microscopy), and hyperspectral datasets, as well as the integration of 18 technical evaluation dimensions covering dataset fidelity, segmentation reliability, model interpretability, and deployment feasibility. To address the widespread lack of explainability, we provide practical guidance for integrating Class Activation Mapping (CAM) and attention heatmaps into industrial inspection workflows, enabling operators to visualize the specific wood-surface regions influencing model decisions. Persistent challenges include the absence of standardized benchmark datasets, limited reproducibility, and minimal adoption of XAI (Explainable AI) techniques. The review concludes by proposing research directions toward more generalizable, interpretable, and industry-ready DL systems for automated wood species identification and quality assessment.</p>