基于Fishmeal-Mask2Former分割模型的掺假鱼粉显微图像识别方法

Detection of Adulterants in Fishmeal Microscopic Images Based on Segmentation Model of Fishmeal-Mask2Former

摘要:

针对传统显微镜检法在鱼粉掺假识别中面临的肉眼辨别难度大、识别效率低的问题，提出一种改进Masked-attention Mask Transformer（Mask2Former）的鱼粉显微图像掺假识别模型，以实现高分辨条件下、复杂鱼粉背景中对异状多目标掺假特征的自动化识别与分割。以鱼粉中掺杂动物源性肉粉为研究对象，构建了Fishmeal掺假鱼粉显微图像数据集，通过形态学分析将掺假特征细分为异常肌肉、骨骼、皮肤、血液和毛发5类组织；开发基于颜色阈值相似性的交互式标注软件，实现逐像素的提示性精确标注；改进Mask2Former分割模型架构，融合ResNet50骨干网络、多头注意力机制和多尺度特征处理机制，增强了鱼粉多样化特征的融合效果；引入重复加权双向特征金字塔网络（Bidirectional feature pyramid network, BiFPN）改进像素解码器，提升了小目标分割能力；引入可学习的尺度级嵌入和掩码注意力（Masked Attention）模块，通过限制交叉注意力关注范围提高了模型细节表现力；在Multi-head Attention层后加入dropout操作防止过拟合。实验结果表明，改进后的Fishmeal-Mask2Former在样本真伪判别阶段整体分类准确率达到98.56%，在精细分割阶段对异常掺假特征定性识别平均精确率达到80.52%、召回率为76.01%、F1分数为78.86%，分割模型平均准确度提升至82.08%，较传统方法具有显著优势。此外，设计分割结果可视化界面，为鱼粉品质检测中的掺假检测环节提供了一种直观、准确与高效的显微视觉自动化识别方法。

Abstract:

Fishmeal, a high-value and necessary protein feed raw material, confronts difficulties in assessing authenticity. The conventional identification process is rendered challenging due to the sophistication of visual discrimination and the inefficacy of microscopic testing techniques by inspector, particularly in the presence of subtle and multiple adulterants. These issues were tackled by proposing an end-to-end segmentation model based on the Masked-attention Mask Transformer (Mask2Former) architecture. Focused on identifying multi-target adulteration feature in complex fishmeal backgrounds under high-resolution conditions to realize automatic recognition and segmentation. Fishmeal was used and adulterated with animal-derived meat powder as an example, an adulterated fishmeal microscopic image dataset was created and the adulteration characteristics were classified into five types of tissues: abnormal muscle, bone, skin, blood, and hair based on morphological analysis. Then an interactive annotation program that utilized color threshold similarity to provide pixel-by-pixel suggestive annotations was developed. The updated Mask2Former model integrated ResNet50, multi-head Attention mechanisms, and multi-scale feature processing mechanisms to extract global information and multi-resolution features. Bidirectional feature pyramid network (BiFPN) enhanced the pixel decoder, thereby improving the model’s performance in processing small-scale targets. The incorporation of the Masked Attention module limited the scope of cross-attention calculations, enabling the model to more effectively focus on the target areas within the mask. The implementation of learnable scale-level embeddings and the reordering of self-attention and cross-attention within the decoder section bolstered the model’s feature learning capabilities. Furthermore, by adding dropout operations after the multi-head Attention layer, the improved Fishmeal-Mask2Former model achieved the overall accuracy rate of 98.56% in distinguishing genuine or adulterated samples. In the fine segmentation stage, the mean average precision of abnormal adulteration feature recognition reached 80.52 %, the mean average recall reached 76.01%, the mean F1 score reached 78.86%, and the segmentation mean accuracy was improved to 82.08%. The method proposed had significant advantages over traditional methods. Finally, a visual strategy with its interface of the segmentation results was designed, aimed to provide an intuitive, accurate, and efficient microscopic visual automated identification method for the fishmeal adulteration in its quality detection.