Cross-Attention Guided Dual-Backbone Fusion for Waste Image Classification

Automated waste classification is a critical enabler of intelligent waste management systems, yet single-backbone deep learning models are limited in their ability to simultaneously capture fine-grained local texture cues and broad semantic representations that together characterise visually similar waste categories. This paper proposes a cross-attention guided dual-backbone fusion architecture that processes input images in parallel through MobileNetV2 and EfficientNetB0, and fuses their 512-dimensional feature vectors via a bidirectional scaled dot-product cross-attention module before classification. Unlike simple concatenation, which treats both feature streams symmetrically, the proposed cross-attention mechanism computes input-conditioned attention weights that allow each backbone to selectively query the key-value representation of the other, dynamically emphasising the most discriminative cross-stream features for a given input. A residual connection preserves the original backbone representations while incorporating the cross-backbone contextual signal. Evaluated on the TrashNet benchmark (2,527 images, six classes) with a stratified 70/15/15 train/val/test split, the proposed model achieves 87.43% test accuracy, 87.61% precision, 87.43% recall, and 87.37% F1-score, outperforming the standalone MobileNetV2 baseline (86.91%) and a concatenation fusion ablation (82.46%) across all metrics. The 4.97 percentage-point gain over concatenation fusion isolates and confirms the contribution of the cross-attention module beyond simple feature combination.