In the context of a continually growing global population and rising food demand, fruit, as an important source of nutrition, requires quality grading and efficient processing, which are critical links in the agricultural supply chain. Fruit grading refers to the classification of fruits based on indicators such as external defects and ripeness, directly influencing their market value and food safety. However, traditional grading has long relied on manual visual assessment, which is not only time-consuming and labor-intensive but also prone to high error rates, making it difficult to meet large-scale processing demands. So, how can more intelligent technologies replace manual labor to achieve automated and precise fruit grading?

Recently, Dr. Muhammad Waqar Akram and his team from the Department of Farm Machinery and Power at University of Agriculture Faisalabad in Pakistan developed a “Machine Vision-Based Automatic Fruit Grading System”, offering a new solution. The related article has been published in Frontiers of Agricultural Science and Engineering (DOI: 10.15302/J-FASE-2023532).

This system deeply integrates “machine vision” and “deep learning”, constructing a fully automated process from defect detection to mechanical sorting. In simple terms, it operates like taking “photos” of fruits, analyzing the details in the images to assess quality, and directing a robotic arm to sort the fruits into different graded boxes. Specifically, the system comprises two key components: a defect detection module and a mechanical sorting module.

Defect detection is the first step in grading. The research employs a “dual-track technical approach”: on one hand, traditional image processing algorithms are used to calculate the proportion of defects on the fruit’s surface through steps such as image preprocessing, threshold segmentation, and morphological operations; on the other hand, convolutional neural networks (CNNs), which excel in image recognition, are introduced to train on image data of mangoes and tomatoes. The dataset used for training the CNN includes both publicly available fruit images and actual captured samples, covering various states from fresh to rotten, ensuring the model can adapt to the complex variations of real-world scenarios. Experiments show that the traditional image processing algorithm achieves detection accuracies of 89% and 92% for mangoes and tomatoes, respectively, while the CNN model demonstrates even higher validation accuracies of 95% and 93.5%, enhancing the reliability of defect identification.

Once the detection results are confirmed, the system sends commands to the mechanical sorting module via a microcontroller (Arduino Uno). This module consists of a conveyor belt and a servo motor-driven sorting arm: as the fruit moves along the conveyor belt, a camera captures images and analyzes them in a designated area. If defects are detected, the sorting arm accurately acts to place the problematic fruit into the corresponding grading box. The entire process is interconnected, taking only a few seconds from image capture to sorting completion, and the hardware costs are relatively low, making it suitable for farms or small processing plants.

It is noteworthy that this system not only addresses the efficiency and error issues of manual grading but also showcases the complementary advantages of “traditional algorithms + deep learning” through technological integration: traditional image processing is fast and cost-effective, making it suitable for scenarios with high real-time requirements; deep learning, on the other hand, can capture more subtle defect features, improving accuracy in complex situations. The research found that even when faced with challenges such as significant color variations on mango skins and complex surface textures on tomatoes, the system remained stable, providing a generalized solution for grading various fruit types.

Currently, the system has demonstrated good practicality in the grading of mangoes and tomatoes. In the future, further optimization of hardware design (such as adding multi-angle cameras) and expanding the range of applicable fruit types could explore its integration with more agricultural scenarios.
DOI: 10.15302/J-FASE-2023532