| 基于深度学习的移动端缺陷蛋检测系统研究 |
| |
| 引用本文： | 范维,胡建超,王巧华,汤文权.基于深度学习的移动端缺陷蛋检测系统研究[J].农业机械学报,2023,54(3):411-420. |
| |
| 作者姓名： | 范维 胡建超 王巧华 汤文权 |
| |
| 作者单位： | 华中农业大学 |
| |
| 基金项目： | 国家自然科学基金面上项目（32072302、31871863）和扬州市科技计划项目（YZ2020047） |
| |
| 摘 要： | 针对缺陷鸡蛋差异性大、人工检测主观性强、实时性差，消费者存在食品安全隐患等问题，提出一种基于深度学习的移动端缺陷蛋无损检测系统，实现对裂纹蛋和血斑蛋的实时检测。首先，建立改进的轻量级卷积神经网络MobileNetV2＿CA模型，以MobileNetV2原网络为基础，通过嵌入坐标注意力机制、调整宽度因子、迁移学习等操作对其进行优化，并进行PC端检测对比试验。试验结果表明：建立的MobileNetV2＿CA模型验证集准确率达93.93%,召回率为94.73%,单个鸡蛋平均检测时间为9.9 ms,对比改进前MobileNetV2模型准确率提升3.60个百分点、召回率提升4.30个百分点、检测时间缩短2.62 ms; MobileNetV2＿CA模型的参数量为2.36×106,较原MobileNetV2网络模型降低31.59%。然后，利用NCNN深度学习框架对MobileNetV2＿CA模型进行训练，并通过格式转换部署至Android移动端，进行NCNN深度学习训练模型的移动端检测验证，及其与TensorFlow Lite深度学习模型的对比分析。试验结果表明：NCNN深度...
|
| 关 键 词： | 缺陷鸡蛋 无损检测 移动端 深度学习 模型优化 |
| 收稿时间： | 2022/5/26 0:00:00 |
Detection System Study of Defective Egg on Mobile Devices Based on Deep Learning |
| |
| FAN Wei,HU Jianchao,WANG Qiaohu,TANG Wenquan.Detection System Study of Defective Egg on Mobile Devices Based on Deep Learning[J].Transactions of the Chinese Society of Agricultural Machinery,2023,54(3):411-420. |
| |
| Authors: | FAN Wei HU Jianchao WANG Qiaohu TANG Wenquan |
| |
| Institution: | Huazhong Agricultural University |
| |
| Abstract: | Aiming at the problems of large diversity of defective eggs, as well as the strong subjectivity and poor real-time detection of artificial detection, and the potential risk of food safety for end-consumers, a non-destructive testing system based on deep learning for defective eggs on mobile device was proposed to realize real-time detection of cracked eggs and bloody eggs. An improved lightweight convolutional neural network MobileNetV2_CA model was firstly established. MobileNetV2 network was taken as the original framework, it was further optimized by embedding coordinate attention mechanism, adjusting width factor, transfer learning and other parameters. The PC detection was also performed for comparison. Results showed that the MobileNetV2_CA model presented the validation accuracy of 93.93%, the recall rate of 94.73%, and the average detection time of 9.9ms for a single egg, which was 3.60 percentage points higher, 4.30 percentage points higher, and 2.62ms shorter than the original MobileNetV2 model, respectively. The parameter score of MobileNetV2_CA model was only 2.36×106, which was 31.59% lower than the original MobileNetV2 network model. In addition, the NCNN deep learning framework was used to train MobileNetV2_CA model, which was further applied to Android mobile terminal through format conversion. The verification of mobile terminal detection of NCNN deep learning training model was investigated and compared with TensorFlow Lite deep learning model. Results showed that the NCNN deep learning model had an average recognition accuracy of 92.72%, an average detection time of 22.1ms for a single egg, and the library file size of 2.7MB, indicating its better performance than TensorFlow Lite and meeting the requirement of practical applications. The effectiveness of the proposed system based on deep learning was finally demonstrated. |
| |
| Keywords: | defective egg non-destructive detection mobile devices deep learning model optimization |
|
| 点击此处可从《农业机械学报》浏览原始摘要信息 |
| 点击此处可从《农业机械学报》下载免费的PDF全文 |
|
