Frequency-Preserving Multi-Resolution Attention CNN: A Learnable Wavelet-Packet Token-Mixing Framework For Banana Disease Detection

Abstract

The initial stages of banana leaf disease, including small necrotic spot and light chlorosis, can be subtle and difficult to detect from natural variation and may be omitted in multiple convolutional downsampling. A CNN-based banana leaf disease classification system using multi-resolution token-mixing attention module (FPMTMA) is proposed in this paper. FPMTMA is inserted between ResNet-50 stages, and introduces learnable wavelet-packet decomposition, cross-band multi-head token mixing, frequency-aware gating and multi-resolution skip recalibration to keep the pretrained backbone behavior unchanged while changing feature-map dimensions. The results on public banana leaf dataset and field collected dataset revealed the mean accuracy of 95.8% and 89.4% respectively and the mean macro F1 score of 95.2% and 88.6% respectively. The early vs late symptom discrimination results an AUC of 0.912. The tested conditions resulted in less performance degradation when compared with the other baselines used in the cross-dataset evaluation. Compared to the baseline, the module has an increase of 1.2 million parameters and around 2.1 ms of inference latency per image for the GPU. These results suggest that learnable frequency-band separation and multi-resolution feature mixing can improve banana leaf disease classification, while robustness under broader field conditions remains to be evaluated.