Research And Application Design Of Mango Grading Detection Algorithm Based On Deep Learning

Mango is a kind of fruit product with high economic value.The added value of mango can be increased by its classification.Traditional mango grading mainly relies on manual completion,which has the problems of low efficiency,strong subjectivity and high cost.Therefore,this thesis studies the mango hierarchical detection algorithm based on deep learning and develops the corresponding application software.The main contents of the thesis are as follows:(1)This thesis made a mango classification detection data set to test the performance of mainstream target detection algorithms in mango classification detection.In order to improve the real-time performance of mango classification detection algorithm and reduce the manufacturing cost of automatic mango classification detection equipment,YOLOv5s was selected as the basic algorithm and a lightweight improvement scheme was designed.Firstly,the backbone feature extraction network and enhanced feature extraction network of YOLOv5s are lightweight improved.Secondly,CA(Coordinate Attention)mechanism was introduced to improve the adaptive attention of network model to mango surface features.In addition,in order to locate boundary boxes more accurately,this thesis studied the application of EIOU(Efficient Intersection over Union)loss function in mango classification detection algorithm,and improved EIOU loss function according to the characteristics of mango classification detection targets.The Complete Intersection over Union(CIOU)loss function in the original algorithm was replaced by the improved EIOU loss function as the positioning loss function.Finally,the effectiveness of the lightweight scheme is verified by the ablation experiment and comparison with other algorithms.In this thesis,the improved lightweight mango grading detection algorithm is called IM＿YOLOv5s.Compared with YOLOv5s,the number of floating-point operations of IM＿YOLOv5s algorithm is reduced by around 77.2%,the number of parameters is reduced by around 72.5%,and the model size is reduced by around 70.8%.In the ablation experiment environment,the reasoning time when batch size is set to 16 is reduced by around38.1%.(2)This thesis designed a desktop mango detection application and a mobile mango detection application.PyQT and Pycharm were used to design a desktop mango classification detection interface.The mobile terminal detection application is designed based on Android architecture.In this thesis,the mango detection application test was completed.The desktop application test showed that the function of mango detection application met the actual use requirements.In the environment with high computing power,the accuracy of IM＿YOLOv5s for mango classification could reach 97.2%.The test on the mobile terminal shows that in the same deployment environment,when the mobile device is combined with the camera for real-time reasoning,the reasoning time of the IM＿YOLOv5s algorithm model is reduced by around 53.1% compared with the original algorithm model from around 584.7ms to around 274.2ms,and the accuracy of mango classification is still 87.2%.