Answers are grounded in this paper and cite their sources.
Convolutional neural networks (CNNs) capture long-range spatial context only indirectly, whereas vision transformers need large training sets. We ask whether adaptively fusing local convolutional and global self-attention features at several depths of a pre-trained CNN improves accuracy and cross-domain robustness. We present HTA-ConvNet, which inserts a Dual-Branch Spatial Attention (DBSA) block after three intermediate stages of EfficientNet-B4. Each block processes the same feature map with a depthwise separable convolution branch and a multi-head self-attention branch and fuses them through a learned channel-wise convex gate with a residual connection. All baselines, including hierarchical vision transformers (Swin, Focal) and hybrid CNN-transformer models (CoAtNet, CvT, ConvNeXt, EdgeNeXt), are re-trained under one protocol. On PlantVillage (54,309 images, 38 classes), HTA-ConvNet attains 99.61% accuracy and 99.58% macro-F1. Because this benchmark is near-saturated, the main evidence comes from near-duplicate-controlled splits, multi-seed confidence intervals, and field data: zero-shot accuracy is 88.34% on PlantDoc (macro-averaged) and 82.47% on Plant Pathology 2020, and we quantify how far fine-tuning and domain adaptation close the remaining gap. Efficient-attention variants, uncertainty-based rejection, and error analyses quantify the cost, reliability, and failure modes of the design.
No takes yet. Share an insight, caveat, or question.
Ramzan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: