Design Of Automatic Fruit And Vegetable Recognition System Based On Machine Visio

Today,with the highly developed informatization and intelligence,people’s lives are also expected to move towards a simpler and more convenient direction.At present,the most common supermarket weighing system is called the weighing and coding integrated scale,which has high labor costs and long time consumption,and is not suitable for scenarios with high pedestrian traffic.Some advanced customer semi self-service smart electronic scales have high operational requirements for customers.Therefore,designing a simple,convenient,and highly automated fruit and vegetable weighing integrated scale meets the requirements of the current intelligent and information society development,and has high practical value.In recent years,machine vision has developed rapidly,and its applications have become increasingly widespread due to its advantages of high accuracy and speed.However,due to computational costs and resources,it is not suitable for small self-service weighing and settlement systems.Studying a lightweight,fast,and high-precision neural network structure can help improve this phenomenon.This thesis focuses on the lightweight design of YOLOv5 network and constructs a lightweight classification algorithm model suitable for real-time monitoring of fruits and vegetables,And conduct hardware construction for verification testing.Firstly,for the recognition and classification scenarios of fruits and vegetables,a self-made dataset was used to collect 3546 images of 28 types of fruits and vegetables,including multiple types of lighting,packaging,and stacking images.To improve the insufficient image quality of certain fruit and vegetable data,geometric transformations and image enhancement were used to expand the dataset,resulting in 24822 different data images.The self-made dataset was used,The lightweight networks ShuffleNetv2 and MobileNetv3 are used to replace the backbone network part of YOOv5 to build a lightweight network and accelerate the feature extraction processing.The highest accuracy can reach 98.27%,and the image recognition speed is only 0.08 s.Secondly,the network model based on MobileNetv3 replacing YOLOv5 backbone is improved,mainly including the following improvement schemes: firstly,for cases with a large variety of fruits and vegetables and varying sizes,such as watermelon and hawthorn,a pre detection head is added to detect targets of different sizes,SPPF is added to the backend of the network backbone to suppress information Loss,and SENet attention mechanism is added to the network to strengthen key features,using ReLU6 and h＿swish’s joint use of swish instead of the original swish activation function can effectively avoid overfitting and gradient disappearance,and reduce the number of parameters while maintaining the same accuracy.Replace IoU Loss with Focal＿EIoU Loss to monitor the model.Finally,in order to improve convergence speed and classification performance,an improved decoupling head was used to replace the coupling head.The experimental results showed that the improved network model reduced the number of parameters by85%,accuracy by only 0.03%,and single image recognition speed by 28.7%.This result can meet the current requirements for fruit and vegetable recognition and classification,and the requirements of the lightweight network model studied in this thesis are also basically implemented.Finally,design the hardware architecture and software system for the fruit and vegetable self-service recognition system,design the customer and administrator system,use Python and Qt interfaces for visualization,and use the STC89C52 microcontroller to collect sensor information and communicate with the upper computer to achieve preset functions.