A hierarchical fusion framework combining wavelet and deep features for accurate recognition of three Hongmu species and visually similar wood species

Due to the high visual similarity in macroscopic textures between Hongmu and its lookalike species, even CNN-based identification methods often struggle to achieve high accuracy. To address the limitations of standard deep learning models in capturing fine-grained textural details, we propose a novel hierarchical fusion framework that integrates wavelet-based high-frequency features into different depths of the Inception-V3 network. Specifically, four fusion strategies were explored to identify the optimal feature integration timing: shallow-layer fusion (after Stem), mid-layer fusion (after Inception-A), deep-layer fusion (after Inception-B), and end-layer fusion (before the classifier). Experimental evidence underscores that the placement of fusion operations critically impacts model effectiveness. Relatively, the end-layer fusion strategy emerged as the optimal architecture, achieving a top-1 accuracy of 93.89%, a 2.39% point higher accuracy than the Inception-V3 baseline. Crucially, this method significantly improved the recognition of highly similar species of Baphia nitida and Pterocarpus tinctorius by effectively leveraging complementary spectral and semantic features. These findings highlight the importance of precise fusion timing for effective multi-scale feature integration. The proposed framework provides a potential efficient and reliable solution for identifying the selected Hongmu species, with significant potential for timber trade oversight and forest conservation.