基于MobileViT-CBAM的枇杷表面缺陷检测方法

Detection Method for Loquat Surface Defect Based on MobileViT-CBAM Network

摘要:

为实现枇杷采后快速、准确筛选，本文以MobileViT为主干特征提取网络，通过分别在Layer1和Layer2层之后嵌入注意力模块CBAM（Convolutional block attention module），强化网络在通道和空间上对细节特征的提取能力，构建了一种轻量化网络模型MobileViT-CBAM。相较于MobileViT，在验证集和测试集上本文方法对疤痕、机械伤、腐烂等缺陷果的识别准确率分别提高1.17、1.23个百分点。试验结果表明，MobileViT-CBAM模型与VGG16、ResNet34、MobileNetV2相比较，准确率最高（97.86%），同时兼具内存占用量小（3.768 MB）、推理时间短（每幅图像需42 ms）的优势。该轻量化网络模型可部署于嵌入式系统。本研究为构建枇杷在线检测系统提供了缺陷识别理论基础，为枇杷等农产品外部品质检测提供了一个高效、准确的方法。

关键词:枇杷;MobileViT-CBAM;缺陷检测;轻量化

Abstract:

The MobileViT as the main feature extraction network was employed in order to accomplish quick and precise post-harvest screening of loquats in the paper. A lightweight network model called MobileViT-CBAM was developed as a result of strengthening the network’s capacity to extract detailed features in both channel and spatial dimensions by inserting convolutional block attention module (CBAM) after Layer1 and Layer2. The method outperformed MobileViT in terms of defect recognition accuracy, showing gains of 1.17 percentage points on the validation set and 1.23 percentage points on the test set for things like scars, mechanical damage, and decaying fruits. According to experimental results, the MobileViT-CBAM model performed better in terms of accuracy (97.86%) than VGG16, ResNet34, and MobileNetV2. It also had the advantage of having a small memory footprint (3.768 MB) and a rapid inference time (42 ms per image). It was possible to use this lightweight network model on embedded systems. The research offered an effective and precise technique for external quality inspection of loquats and other agricultural products by providing a theoretical framework for fault recognition in the construction of an online detection system for loquats.

Key words:loquat;MobileViT-CBAM;defect detection;lightweight