Answers are grounded in this paper and cite their sources.
s In the cultivation mode of mushroom logs, detecting shiitake mushroom fruiting bodies is a crucial step toward enabling automated harvesting and factory-based production. However, existing methods still face several limitations. Most models rely solely on Red, Green, Blue (RGB) images and lack spatial depth perception, which leads to inaccurate detection of mushrooms with varying growth heights. The texture similarity between mushroom caps and cultivation logs makes it difficult to extract fine-grained features in complex backgrounds. In addition, immature mushrooms, which are often small and densely distributed, are easily missed during downsampling operations. To address these challenges, this paper proposes a Multi-modal and Detail-aware Fusion DEtection TRansformer model (MDF-DETR). MDF-DETR enhances shiitake mushroom detection by incorporating monocular depth estimation to construct Red, Green, Blue and Depth (RGB-D) inputs, which improves spatial perception of mushroom positions and height differences. A Light Residual Attention (LRA) block is embedded into the redesigned C2fL module to strengthen fine-grained texture representation under complex backgrounds. Additionally, a Small Object Feature Fusion Network (SOFN), which combines Space-to-Depth Convolution (SPDConv) and a Cross-Stage Partial Omni-Kernel (CSPOmniKernel) module, is introduced to better retain detailed features and improve the detection of small and dense mushroom targets. The model was trained and evaluated on a self-collected dataset. This dataset was captured by a fixed-position overhead camera and contains 1000 annotated RGB images of shiitake mushroom logs. Depth maps were generated through monocular depth estimation and concatenated with the RGB images to form RGB-D inputs. Experimental results show that MDF-DETR achieves a precision of 90.0 %, a recall of 85.4 %, and a mean average precision (mAP) of 90.3 %, with 14.1 million parameters and 57.0 GFLOPs. These results confirm that MDF-DETR not only outperforms the baseline RT-DETR model but also achieves a good balance between detection accuracy and computational complexity.
No takes yet. Share an insight, caveat, or question.
Wang et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: