Scholarlink

<p>Foodborne bacteria stand as the foremost cause of human mortality, representing the deadliest disease in existence. Cutting-edge research in the field of foodborne bacteria detection leverages the power of deep learning and ensemble learning techniques, however no one exclusively focusing on feature fusion to distinguish between live and dead bacterial cells, ensuring food safety. This paper introduces a novel feature fusion approach with improved naïve inception module, designed to enhance the classification of live and dead foodborne bacteria using hyperspectral microscope imagery dataset. The approach combines the strengths of Efficient Networks, specifically <i>EfficientNet-B1</i>, and the established architecture of <i>ResNet-50</i>. We have enhanced the original naïve inception module by substituting the conventional activation function with LeakyReLU, a modified activation function. One notable advantage of LeakyReLU is that it helps mitigate the vanishing gradient problem, which can lead to more stable and efficient training of deep neural networks. The modified inception module excels at capturing contextual information, enhancing its ability to recognize intricate patterns within images. By fusion of these models via modified inception module, enriching feature representations encompassing image characteristics. Following feature extraction from the integrated <i>EfficientNet-B1</i>, and <i>ResNet-50</i> models, the subsequent phase of our methodology involves transmitting these fused features to a naïve inception <i>module</i> for conclusive classification between live and dead bacterial entities. Our dataset comprises a total of 1274 images, with 656 images corresponding to dead cells and 618 images representing live cells. We applied histograms to enhance the image quality by providing insights into the distribution of pixel intensities, aiding in contrast adjustment and overall image enhancement. Following rigorous preprocessing steps, Hyperparameter optimization is crucial for enhancing deep learning classifier performance, as it fine-tunes model settings, leading to superior outcomes like a 90.90% accuracy rate. Additionally, conducting an ablation study helps analyze the impact of modifications on a model, providing valuable insights for further improvements. Foodborne bacteria prediction empowers food safety departments with early identification of potential outbreaks, enabling timely intervention and proactive measures to ensure food safety.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Advancing food safety: deep learning for accurate detection of bacterial contaminants

  • 期刊论文

  • 发表日期：

    2026年1月21日

查看全文
  • Saif Ur Rehman Khan,
  • Omair Bilal,
  • Arash Hekmat,
  • Inzamam Shahzad,
  • Asif Raza

摘要

Foodborne bacteria stand as the foremost cause of human mortality, representing the deadliest disease in existence. Cutting-edge research in the field of foodborne bacteria detection leverages the power of deep learning and ensemble learning techniques, however no one exclusively focusing on feature fusion to distinguish between live and dead bacterial cells, ensuring food safety. This paper introduces a novel feature fusion approach with improved naïve inception module, designed to enhance the classification of live and dead foodborne bacteria using hyperspectral microscope imagery dataset. The approach combines the strengths of Efficient Networks, specifically EfficientNet-B1, and the established architecture of ResNet-50. We have enhanced the original naïve inception module by substituting the conventional activation function with LeakyReLU, a modified activation function. One notable advantage of LeakyReLU is that it helps mitigate the vanishing gradient problem, which can lead to more stable and efficient training of deep neural networks. The modified inception module excels at capturing contextual information, enhancing its ability to recognize intricate patterns within images. By fusion of these models via modified inception module, enriching feature representations encompassing image characteristics. Following feature extraction from the integrated EfficientNet-B1, and ResNet-50 models, the subsequent phase of our methodology involves transmitting these fused features to a naïve inception module for conclusive classification between live and dead bacterial entities. Our dataset comprises a total of 1274 images, with 656 images corresponding to dead cells and 618 images representing live cells. We applied histograms to enhance the image quality by providing insights into the distribution of pixel intensities, aiding in contrast adjustment and overall image enhancement. Following rigorous preprocessing steps, Hyperparameter optimization is crucial for enhancing deep learning classifier performance, as it fine-tunes model settings, leading to superior outcomes like a 90.90% accuracy rate. Additionally, conducting an ablation study helps analyze the impact of modifications on a model, providing valuable insights for further improvements. Foodborne bacteria prediction empowers food safety departments with early identification of potential outbreaks, enabling timely intervention and proactive measures to ensure food safety.