【摘要】 为实现果面缺陷冬枣实时检测，并解决缺陷的尺寸与位置不同影响检测精度的问题，提出一种基于改进单次多目标检测器(Single shot multibox detector, SSD)的果面缺陷冬枣实时检测方法。以陕西大荔冬枣中的虫蛀、轮纹和木质化3种缺陷果和正常果为研究对象，在数据采集设备下采集实际分拣图像，然后通过数据增强由400张扩充至2 000张。改进SSD,建立MobileNetV3-SSD模型，为实时检测奠定基础；引入改进感受野块(RFB)可实现模型多尺寸提取冬枣缺陷特征的能力；用空间注意力模块(SAM)代替挤压和激励通道注意力模块(SE)增强模型定位冬枣缺陷特征的能力。试验结果表明，本研究模型在果面缺陷冬枣数据集上的表现均优于目前先进目标检测网络模型(RetinaNet和EfficientDet-D0),该模型对4类冬枣的整体检测精准性(mAP)达到91.89%,检测速度达到1 s 40.85帧。因此本研究模型较好地平衡了实时性和精准性，可应用于果面缺陷冬枣分拣流水线。更多还原

【Abstract】 In order to realize the real-time surface defect detection of winter jujube and solve the problems that different sizes and positions affected the detection accuracy, a real-time surface defect detection method of winter jujube based on improved single shot multibox detector(SSD) was proposed. Three kinds of defective winter jujubes(worm, wheel-pattern and lignification) and normal winter jujubes from Dali(Shaanxi province) were taken as the research objects. The actual sorting images were collected by data acquisition equipment, and then expanded from 400 to 2 000 by data enhancement. The SSD was improved, and MobileNetV3-SSD model was established to lay the foundation for real-time detection. The introduction of improved receptive field block(RFB) could realize the ability of model to extract the defect feature of winter jujube at multiple scales. Spatial attention module(SAM) was used to replace squeeze-and-excitation(SE) block, so the ability of the model to locate the defect feature of winter jujube was enhanced. The test results showed that the performance of the proposed model on the dataset of defective winter jujube was better than the current advanced target detection network models(RetinaNet and EfficientDet-D0). The averall detection accuracy of the model for four types of winter jujube was 91.89%, and the detection speed was 40.85 frames per second. Therefore, the model established in this study can balance the real-time performance and accuracy, and can be applied to sorting pipeline of winter jujube with surface defect.更多还原