基于机器视觉与深度学习的猴头菇质量检测方法

Quality Detection Technology of Hericium erinaceus Based on Machine Vision

摘要:

猴头菇作为我国传统的食药两用真菌,其商品化过程中对质量的快速、客观评估提出了迫切需求。针对传统人工感官判定方法效率低、标准不一的问题,提出了一种基于机器视觉与深度学习的猴头菇质量智能检测方法。首先,构建了一种多尺度优化的分水岭图像分割算法,以精准提取猴头菇边界并抑制复杂背景下的过分割现象。随后,通过图像增强与数据增强策略提升图像质量与模型鲁棒性。在分类阶段,引入Swin TransformerTiny网络结构,结合局部窗口注意力与移位窗口机制,实现对猴头菇正常与异常样本的高精度判别。实验结果表明,所提出方法在实际采集图像上取得了94.3%的分类准确率,相较于ResNet-18、EfficientNet、DilateFormer、MambaVision等主流模型分别提高了3.5、2.2、1.5、1.2个百分点,具备良好的工程实用性与部署价值。

Abstract:

Hericium erinaceus is a traditional edible and medicinal fungus with high nutritional and economic value, and its commercial circulation requires rapid, objective, and stable quality evaluation. However, manual sensory grading is easily affected by evaluator experience, illumination conditions, and fatigue, which leads to low efficiency and inconsistent judgment results in large-scale production scenarios. Aiming to address these problems, an intelligent quality detection method for H. erinaceus based on machine vision and deep learning was proposed. Firstly, images of normal and defective samples were collected under practical acquisition conditions, and preprocessing operations were used to reduce background interference and improve the visibility of surface texture features. A multi-scale optimized watershed segmentation algorithm was then constructed to extract the target region accurately, suppress over-segmentation in complex backgrounds, and provide reliable input for subsequent feature learning. On this basis, image enhancement and data augmentation strategies were introduced to enrich sample diversity, alleviate the influence of limited datasets, and strengthen the generalization ability of the model. In the classification stage, a Swin Transformer Tiny network was adopted to combine local window attention with shifted window attention, so that both fine-grained local texture information and broader contextual relationships can be effectively captured. The model was used to discriminate between normal and defective H. erinaceus samples, and its performance was compared with several mainstream deep learning models under the same experimental conditions. Experimental results showed that the proposed method achieved a classification accuracy of 94. 3% on real-world collected images, outperforming ResNet-18, EfficientNet, DilateFormer, and MambaVision by 3. 5, 2. 2, 1. 5, and 1. 2 percentage points, respectively. The results indicated that the combination of accurate segmentation and attention-based feature extraction can improve the reliability of H. erinaceus quality detection. The proposed system demonstrated strong engineering practicability and deployment potential, providing a feasible technical route for intelligent quality inspection of structurally complex edible fungi and supporting the standardization and intelligent development of the edible and medicinal fungus industry.