A Multiple Attentions-Based Multilevel Hybrid-Guided Deep Fuzzy Convolutional Neural Network for Image Recognition

In recent years, as a vital branch of artificial intelligence, image recognition (IR) has garnered significant attention across various industries. Although deep convolutional neural networks and attention-based models have achieved remarkable advancements in this field, they each exhibit inherent limitations and struggle to effectively handle uncertainty. To address these challenges, this article proposes a multiple attentions-based multilevel hybrid-guided deep fuzzy convolutional neural network (MAMH-DFCNN), aimed at enhancing recognition accuracy and uncertainty processing ability while simultaneously expanding the application scope of deep fuzzy systems. Specifically, by introducing a novel squeeze-and-excitation (SE) attention mechanism-based interval type-2 generalized fuzzy hyperbolic tangent autoencoder, a multilevel convolutional fuzzy feature extraction framework is developed to capture features across different spatial and granularity levels, enriching feature representations and improving the model’s robustness against uncertainty. A spatial pyramid-type fuzzy feature attention fusion module (FFAFM) is then employed to fully integrate the multilevel fuzzy features, acquiring more abstract semantic information while reducing information redundancy. Subsequently, a hybrid-driven cross-attention classifier is designed to facilitate the deep interaction of features with different granularities, ultimately producing the final recognition outcome. Extensive comparative experiments and analytical evaluations are conducted. The results demonstrate that MAMH-DFCNN achieves desirable performance across both binary and multiclass classification tasks, exhibiting certain competitiveness against baseline models and state-of-the-art models. These findings highlight the feasibility and effectiveness of MAMH-DFCNN in improving recognition accuracy and robustness against uncertainty in IR.