Grading Machine Innovation Challange

Grading size and Quality of Local Tomatos cultivated in Eastern Hills of Nepal

Grading_machine

Project Information

Project Overview

This project was developed in collaboration with the support of swisscontact(Sahaj) Project and DELTA 3.0 at the Institute of Engineering, Purwanchal Campus. The Project aims to grade the spherical vegetables(Tomatoes). Grading is multistepped process where Size, and Quality are to be graded.The system combines computer vision and mechanical systems to automate the categorization of vegetables based on their size and ripeness.

Methodology

The Grading machine is roughly divided in 2 modules or category

  • Size Grading
    • This is first and primary stage of grading. in this stage the tomatoes are graded on the basis of their sizes. The hopper had been desiged for ensuring the height as well control of tomato to grader. Gravity seperation with slope and divergence is applied. multiple guides with gradual increase in seperation coupled with optimal slope that allows tomato to fall at controlled rate would segregate the tomatoes on based of size. 3 Sizes had been standarized after market studying and their demand in different industry from ketchup, salad to normal consumptions.

    These size grading proved effective means to segrgrate the farmers collection in the collection centre of these region in Eastern Hills, However The grading often have to come with the quality. for quality , the ripeness is to be determined and for this purpose Computer Vision had been used as premodel of Cyber-physical System in Agriculture production

  • Quality Grading with Machine Vision & Learning
    • For Quality Grading, The Different approach had been used. Two of most advanced techniques had been used namely Machine Vision & Machine Learning The samples are captured with the USB Camera from Raspberry-Pi. samples are then annotated with reflection, rotation, Zoom, and shearing to generate the enough number of images for the Machine Learning algorithm to be deployed.
    Features
    Defective Sample image
    Localization Results
    Riped Sample image
    Localization Results
    unriped Sample image

    The images then used to train a network that leverages the Transfer Learning from MobileNet V2 which shall help train the network on the foundations of images followed by average pooling. The base model thus prepared is coupled with dense layers of neural neteorks of 512, 1024, 512 neurons consecutively of each having activation function as ReLU and optimizer being Adam. The network was trained for 10 epoch obtaining accuracy of 0.975 (Normalized to 1) and thus obtained model had been saved for deployment.

    Localization Results
    Accuracy & Losses over the epochs

    Results

    Features
    Ripe with Masks
    Localization Results
    Unripe with Masks

    The model was then used along with implementing RGB Masks highlighting areas after as well as categorizing the tomato quality. The model was deployed in cloud whereas real time picture is captured by the camera and fed to network leveraging the computational offloading and provide results.

    Conclusion

    This Grading machine was first prototype that requires lots of iteration. key takeouts are quality and Size Grading coupled with Machine Vision & Edge computing. However improvements can be made in all the sectors as

  • Automation in quality and size grading as integrated unit
  • Deployment of model at Embeddeds and Model optimization for more real time operations.