Machine Vision-Based Detection of Browning Maturity in Shiitake cultivation sticks

Zeting Liu¹Jiuxiao Zhao^2,3Wengang Zheng^3,4Qiuxiao Song⁴Xin Zhang^3,5Wei Liu²Feifei Shan^2,3Ruixue Xu²Zuolin Li^2,6Jing Dong^2,6Pengfei Zhao^2,6Yajun Wang^1*Mingfei Wang^2,3*

• ¹Dalian Polytechnic University School of Mechanical Engineering and Automation, Dalian, China
• ²NongXin Science & Technology (Beijing), Beijing, China
• ³Intelligent Equipment Technology Research Center, Beijing Academy of Agriculture and Forestry Sciences, Beijing, China
• ⁴Shandong Qihe Bio Technology, Shandong, China
• ⁵Beijing Academy of Agriculture and Forestry Sciences, Beijing, China
• ⁶Beijing Academy of Agriculture and Forestry Sciences Information Technology Research Center, Beijing, China

Accurate monitoring of pigmentation changes during the browning stage of shiitake cultivation sticks is essential for assessing substrate maturity, forecasting mushroom emergence, and improving cultivation quality. However, current commercial detection methods lack objective, real-time and quantifiable evaluation indicators for the browning degree assessment. To address this, this study proposes a detection method based on two-stage image segmentation. Firstly, the proposed VG-Stick-YOLOv11, built upon YOLOv11n-seg with VanillaNetBlock and GhostConv, achieved the best mIoU of 95.80% while markedly reducing parameters and computation, offering an efficient solution for real-time contour extraction and browning stage classification of shiitake sticks. Based on the extracted features during this stage, machine learning techniques were employed to enable rapid and semiautomatic annotation of browning regions, thereby constructing a segmentation dataset. Finally, the ResNet-Stick-UNet (RS-UNet) model was designed for browning regions segmentation and area ratio calculation. The encoder adopts ResNet50 with multi-branch inputs and stacked small kernels to enhance feature extraction, while the decoder incorporates a hybrid structure of grouped and depthwise separable convolutions for efficient channel fusion and detail preservation. In addition, a spatial attention mechanism was embedded in skip connections to emphasize large-scale browning regions. Experimental results showed that RS-UNet achieved a segmentation accuracy of 94.35%, improved the IoU to 88.56%, outperforming comparison models such as Deeplabv3+ and Swin-UNet, while reducing parameters by 36.31% compared to ResNet50-U-Net. This collaborative model provides a quantitative basis for the maturity detection of shiitake cultivation sticks during the browning stage, promoting the intelligent and standardized development of shiitake substrate cultivation.