DDHTS-Net: dual-domain hierarchical texture supervision network for plant texture analysis

Texture is a fundamental low-level visual attribute characterized by complex local patterns and spatial structures. Plant textures, in particular, exhibit subtle inter-class differences and substantial intra-class variability, making accurate recognition challenging. This difficulty is further increased by fine-grained local texture variations and imaging conditions such as scale changes, illumination variations, and viewpoint differences. Existing deep texture analysis models often rely on a single image domain or representation stream, limiting their ability to capture complementary cross-frequency cues and hierarchical abstraction differences required for fine-grained texture discrimination.

To address this challenge, this study proposes a dual-domain hierarchical texture supervision network, DDHTS-Net, designed to capture and model complex texture attributes across multiple domains and hierarchical levels. In DDHTS-Net, four channels with different granularity levels are constructed using the Undecimated Wavelet Transform. The network integrates an intra-channel hierarchical supervision unit and an inter-channel granularity-level supervision unit. The intra-channel unit treats deep feature maps from different layers as experts with distinct levels of abstract knowledge, enabling higher-level experts to guide lower-level counterparts in discovering latent and critical texture attributes. The inter-channel unit uses experts at different granularity levels within the same abstraction hierarchy to capture distinct spatial frequency characteristics and further supervise the latent texture features in each layer of the original image channel.

Extensive experiments on two plant classification datasets and two benchmark texture datasets demonstrate that DDHTS-Net consistently achieves superior classification performance compared with existing state-of-the-art methods.

These results indicate that integrating dual-domain representations with hierarchical texture supervision can effectively enhance fine-grained texture discrimination. The proposed DDHTS-Net provides an effective framework for modeling complex texture attributes in plant recognition and general texture classification tasks.