Non-destructive Detection of Soybean Appearance Quality Based on Improved RetinaNet
Quickly,accurately,and effectively distinguishing the appearance quality of soybeans is an important and arduous task in soybean food quality inspection,food safety,and packaging.In this article,a soybean appear-ance quality detection model based on an improved convolutional neural network RetinaNet was proposed.Replacing the original backbone network ResNet50 with ResNet34 reduced the number of model parameters,improved computa-tional speed,and reduced computational time while ensuring accuracy.Embedding ECA modules at the output ends of the backbone network and feature pyramid(FPN)further extracted advantageous features,reduced the impact of redundant features on the network,and improved network performance.At the same time,in order to ensure the rich-ness of the original features,the output of the ECA module embedded after FPN was overlaid with the output of the backbone network,and the obtained features were used as inputs and passed into the classifier for recognition and de-tection.The results indicated that the improved RetinaNet soybean quality detection model proposed in this article had a precision of 97.39％,an mAP value of 98.64％.
NETL
NSTL
万方数据
