Dual-Branch Cross-Feature Fusion of RGB and Edge Maps for Fine-Grained Electronic Component Recognition

Abstract

Finer-grained recognition of visually similar electronic objects is not that easy since the category distinctions can be based upon subtle appearance and boundary information. This work has introduced a dual-stream RGB–edge classification architecture that jointly finds semantic representations via a pretrained ConvNeXt-Tiny backbone with structural representations via Canny edge maps via a lightweight convolutional branch. The two streams of features are combined with learnable feature gating and are classified by a fully connected prediction head. A stratified train, validation and test partitions were used to run experiments on a balanced dataset of 950 images of 10 categories of electronic objects. Across three random seeds, the proposed model achieved an accuracy of 94.64% ± 0.87, a Macro F1-score of 94.55% ± 0.87, and Cohen’s kappa of 0.940 ± 0.010. In the reported experiments, the dual-stream model had a competitive performance even though the RGB-only ConvNeXt-Tiny baseline achieved a higher average accuracy, and the variability was less. These findings suggest that edge features can be used to complement structural information, although this depends on the size of data, quality of edges, and fusion design. More and bigger datasets should be assessed before it can be claimed to be deployment-oriented.