Research on Cherry Defect Detection and Recognition Based on Feature Fusion Attention Mechanism
In view of the existing problems in cherry defect detection and recognition,and to realize in-telligent rapid detection and accurate recognition,a lightweight defect detection and recognition model based on convolutional neural network was proposed for cherry images,which could provide a theoretical basis for developing lossless intelligent detection system carried on mobile terminal.Firstly,the collected images of in-tact cherries,growth-stimulated cherries,twin cherries and rotten cherries were preprocessed,and then were divided into training,validation and test sets in proportion.Secondly,after comparing the network models such as NASNet-Mobile,MobileNetV2,ResNet18,InceptionV3 and VGG-16 based on transfer learning,the Mo-bileNetV2 with good performance in all aspects was selected as the baseline model,and then the I-Mobile-NetV2 model was established after fine tuning.On the basis of I-MobileNetV2,the coordinate attention was embedded,and then the ICA-MobileNetV2 model was constructed.The average accuracy of ICA-MobileNetV2 model reached 97.09% ,which was 7.85% higher than that of baseline model(90.02% )and 2.91% higher than that of I-MobileNetV2 model(94.34% ).As a deployable lightweight model,ICA-MobileNetV2 had high-er accuracy and fewer parameters,so it was suitable for cherry defect detection and multi-classification tasks,which provided a new idea for cherry defect detection and quality classification research.
国家科技期刊平台
NSTL
万方数据
