PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Learning and Machine Vision for Food Processing: A Survey

Lili Zhu, Petros Spachos, Erica Pensini, Konstantinos Plataniotis

arXiv:2103.16106v1cs.CVcs.LG

TL;DR

Food processing requires efficient quality and safety control across complex, often labour-intensive operations. This survey reviews machine-vision systems, imaging techniques, and traditional and deep learning methods used in food processing, along with open challenges and future trends. It reports broad use across safety inspection, monitoring, grading, and foreign-object detection, while identifying environmental and feature-extraction limitations that constrain robustness and reliability.

  • Problem

    Labour-intensive food-processing operations create pressure to reduce costs, improve quality, and increase processing efficiency while maintaining food safety.

  • Method

    The survey synthesizes recent machine-vision techniques and traditional and deep learning methods for food-processing applications.

  • Results

    The reviewed systems are used across food safety inspection, processing monitoring, foreign-object detection, grading, and packaging, with reported examples including 98.13% grain-monitoring accuracy and accuracy > 97.1% for general packaging.

  • Takeaways & Limitations

    Machine vision provides faster and more efficient working methods in food processing and is expected to see increased use despite remaining challenges.

  • Takeaways & Limitations

    MVS performance depends on measurement conditions and environment, while food diversity leaves deficiencies in detecting subtle colour changes and unusual shapes.

Abstract

from arXiv · show

The quality and safety of food is an important issue to the whole society, since it is at the basis of human health, social development and stability. Ensuring food quality and safety is a complex process, and all stages of food processing must be considered, from cultivating, harvesting and storage to preparation and consumption. However, these processes are often labour-intensive. Nowadays, the development of machine vision can greatly assist researchers and industries in improving the efficiency of food processing. As a result, machine vision has been widely used in all aspects of food processing. At the same time, image processing is an important component of machine vision. Image processing can take advantage of machine learning and deep learning models to effectively identify the type and quality of food. Subsequently, follow-up design in the machine vision system can address tasks such as food grading, detecting locations of defective spots or foreign objects, and removing impurities. In this paper, we provide an overview on the traditional machine learning and deep learning methods, as well as the machine vision techniques that can be applied to the field of food processing. We present the current approaches and challenges, and the future trends.

1. Introduction

Food processing affects quality of life, economic development, and food safety, creating a need for accurate and efficient quality control. The survey organizes recent machine-vision and image-processing research to track state-of-the-art methods in food processing.

  • 1. Introduction: Food processing transforms raw animal or plant materials, or existing foods, into forms suited to modern dietary habits.It involves physical and chemical changes while preserving nutritional properties and avoiding toxic or harmful substances.
  • 1. Introduction: Food processing is important to food scientists, industry, and consumers because it is linked to quality of life, economic development, and food safety.
  • 1. Introduction: Quality control of food and agricultural products must be accurate and efficient to meet increasing societal expectations and standards.
  • 1. Introduction: Machine vision and image processing can collect food size, weight, shape, texture, colour, and other details for monitoring and control while reducing repetitive-work errors.
  • 1. Introduction: The survey summarizes representative papers from the last five years and organizes them by machine-vision functions and methods.Its structure covers machine vision systems, food processing, machine learning, open challenges, future trends, and conclusions.

2. Machine Vision System (MVS)

A machine vision system automatically acquires and processes images of stationary or moving objects, then uses the resulting data to control manufacturing procedures. Its core operations are image acquisition and image processing, which support objective, non-destructive food evaluation.

  • 2. Machine Vision System (MVS): A machine vision system uses cameras, illumination, computer-based image-processing software, and mechanical systems to observe, evaluate, and recognize objects.The system can then control subsequent manufacturing procedures.
  • 2. Machine Vision System (MVS): An MVS has two main parts: image acquisition and image processing.Together, they enable objective and non-destructive food evaluation.
  • 2. Machine Vision System (MVS): Image acquisition determines image quality and information, forming the foundation for effective subsequent image processing.
  • 2. Machine Vision System (MVS): Image processing guides machinery operation and supports the multi-tasking of machine vision systems.
  • 2. Machine Vision System (MVS): The survey refers to Table 2 for the main processes of a typical MVS before describing each component in detail.

2.1. Image Acquisition

Image acquisition in food-processing machine vision uses cameras, networked transmission, and specialized imaging modalities to obtain visual, depth, spectral, thermal, chemical, or internal-structure information. These modalities support tasks including measurement, grading, quality assessment, temperature monitoring, and foreign-object detection.

  • 2.1. Image Acquisition: MVSs acquire real-time images through photos, videos, and 3D technologies, transmitting them by cable, Ethernet, radiofrequency identification, or wireless sensor networks.
  • 2.1.1. Stereo systems: Stereo systems estimate object depth and can combine 2D and 3D surface data for food measurement, mass-based grading, and defect detection.Potato applications use depth-derived length, width, thickness, volume, and mass information alongside surface data.
  • 2.1.2. Remote sensing image (RS): Remote sensing images use multiple electromagnetic-wave bands and geographic information, with geometric processing aligning images to ground coordinates.NIR spectroscopy has been used to assess apple soluble-solids content, firmness, and internal flesh browning.
  • 2.1.3. Hyperspectral imaging: Hyperspectral imaging captures image information across many wavelength-specific channels, providing both spatial and spectral data.Applications include detecting peach-surface fungi and assessing fruit quality or egg freshness and defects.
  • 2.1.4. X-ray imaging: X-ray imaging detects metal and denser nonmetallic contaminants, including glass, calcified bone, stone, and high-density plastic.
  • 2.1.5. Thermal imaging: Thermal imaging detects infrared heat without contact, converts it into an electrical signal, and generates images for determining object temperature.It can monitor food temperature during pretreatment, pre-heating, and disinfection.
  • 2.1.6. Magnetic resonance imaging (MRI): MRI uses nuclear magnetic resonance to characterize food by tracking proton motion with good spatial resolution.It has been used to determine food texture and address internal heat-transfer, moisture-migration, and temperature-distribution measurements.
  • 2.1. Image Acquisition: Other imaging systems include 0.3 THz orthogonally polarized imaging for detecting foreign objects on moving food conveyors and multimode spectroscopic systems for food safety and quality.

2.2. Image Processing

Image processing transforms and analyzes food images through low-, intermediate-, and high-level stages. These stages improve image quality, isolate targets, describe features, and support recognition for subsequent processing decisions.

  • Image processing generates or improves images to characterize a region of interest without interpreting the image’s content or meaning.It is organized into low-, intermediate-, and high-level processing.
  • Low level processing: Low-level processing pre-processes imperfect images to improve subsequent analysis despite lighting, resolution, distance, and other imaging limitations.Images may be acquired through multiple imaging and sensing devices and converted into digital forms.
  • Intermediate level processing: Intermediate-level processing includes segmentation, representation, and description, with segmentation separating targets from unwanted information to reduce computational cost and improve accuracy.Watershed segmentation imitates geographic features to classify objects, while boundary and regional representations describe characteristics such as size and shape.
  • Intermediate level processing: A sliding comparison local segmentation algorithm achieved 93.8% accuracy distinguishing orange stem ends from peels and 97% accuracy detecting defective samples across 1191 oranges.
  • High level processing: High-level processing uses statistical or deep learning methods for image recognition and interpretation, with results determining subsequent processing requirements.KNN, SVM, neural networks, fuzzy logic, and genetic algorithms can help interpret image information.
  • High level processing: Figure 3 presents image interpretation techniques applied to different image elements, while Table 3 summarizes non-machine-learning classification methods.

3. Food Processing

Food processing spans primary and deep processing while requiring quality, safety, monitoring, packaging, and foreign-object controls. Traditional operations and evaluations create labor, cost, time, and destructiveness challenges that motivate automated and non-destructive approaches.

  • 3.1. Food Processing: Food processing transforms raw materials or existing foods into edible products, with primary processing preserving nutrients or enabling transport and deep processing further improving product characteristics.Examples include drying, shelling, milling, slaughtering, freezing, and producing bread, noodles, or biscuits.
  • 3.1. Food Processing: 17Traditional food-processing operations are labour-intensive, limiting resource optimization and contributing to high labor costs, raw-material waste, and difficulty achieving high quality at low price.
  • 3.1. Food Processing: Quality and safety procedures include evaluation, processing monitoring, packaging, and foreign-object detection to prevent resource waste and improve productivity.Appearance, texture, and food components are primary references for quality and safety assessment.
  • 3.1. Food Processing: Traditional food safety and quality evaluations are often time-consuming and destructive, creating demand for rapid, non-destructive assessment methods.Food safety concerns include contamination, deterioration, toxic substances, and nutritional requirements.
  • 3.1. Food Processing: Size, shape, colour, and texture are used to grade food and evaluate quality or safety because they relate to market price and can reflect intrinsic chemical changes.Image processing can quantitatively characterize size and shape through measurements such as projected area and perimeter.
  • 3.1. Food Processing: Automation and remote monitoring are key for controlling temperature- and pressure-sensitive processing while reducing human error from continuous manual observation.
  • 3.2. Packaging Monitoring: Packaging protects food, supports sanitation, display, and transportation, but non-automated monitoring can cause human error, low efficiency, and undesirable objects entering packages.
  • 3.3. Foreign Object Detection: Foreign objects such as insects, glass, metal, and rubber can enter food during processing, causing consumer harm, lost brand loyalty, recalls, and rejection.Detection by naked eye is difficult, whereas technologies such as X-ray and ultrasound can detect foreign objects.

4. Machine learning approaches

Machine learning discovers patterns from training samples for prediction, while deep learning extends this approach with more complex architectures and greater data-analysis capacity. The paper situates both within the broader relationship among artificial intelligence, machine learning, and deep learning.

  • Machine learning uses training samples to discover patterns and achieve accurate predictions of future data or trends.
  • Deep learning developed in response to the increasing scale and difficulty of data analysis, using more complex architectures and improved data-analyzing capacity than traditional machine learning.
  • Figure 4 depicts the relationship between artificial intelligence, machine learning, and deep learning.

4.1. Traditional Machine Learning Methods

Traditional machine learning methods support classification, clustering, dimensionality reduction, and food-processing evaluation tasks. The survey describes their operating principles and reports applications in grading, quality assessment, monitoring, and foreign-object detection.

  • Learning paradigms: Traditional machine learning extracts features manually and includes supervised, unsupervised, and reinforcement learning according to training data and labels.
  • Supervised learning: SVM classifies samples by finding a hyperplane that maximizes their separation, while converting classification into a convex quadratic programming problem.
  • Neighbour and clustering methods: KNN assigns a sample to the category containing most of its K closest feature-space neighbours, whereas K-means clusters unlabeled continuous data using a prespecified number of categories.
  • Probabilistic methods: Bayesian networks represent uncertain causal reasoning with directed acyclic graphs and factor joint probabilities into local conditional distributions.
  • Dimensionality reduction: PCA and LDA map samples into lower-dimensional spaces, with PCA maximizing mapped-sample divergence and LDA maximizing classification performance.
  • Food-processing applications: Traditional machine learning applications include food grading, monitoring, and detection, with reported accuracies ranging from 89.2% for fuzzy-logic rice monitoring to 100% for Bayesian-network olive-oil grading.

4.2. Deep Learning Methods

Deep learning uses multilayer representational learning for food-processing vision, with CNNs, RNNs, FCNs, ANN, and BP networks supporting recognition, grading, monitoring, packaging, and defect detection. Reported applications achieve high accuracy across diverse food tasks, while FCNs additionally provide pixel-level outputs for variable-sized images.

  • Deep Learning Fundamentals: Deep learning builds models by iterating functions across multiple layers and learns increasingly abstract representations from image data.Its reported learning effectiveness is superior to traditional machine learning, although interpretability is lower.
  • Neural Network Architectures: ANNs use weighted sums across input, hidden, and output layers, whereas BP networks update weights by back-propagating loss gradients.BP training requires a known expected output for each input.
  • Neural Network Architectures: CNNs combine convolutional, nonlinear, pooling, and fully connected layers to map image matrices to category probabilities.Pooling compresses feature-space dimensions while preserving depth, and ReLU returns zero for negative inputs while retaining positive values.
  • Deep Learning Architectures: FCNs replace CNN fully connected layers with convolutional layers, accept images of any size, and produce labelled pictures through up-sampling.This supports semantic segmentation by pixel-level image classification.

5. Open challenges and Future Trends

Machine vision applications in food processing remain constrained by environmental sensitivity, diverse food characteristics, and computational demands. Future work emphasizes more efficient, robust image processing, embedded systems, and integration with other technologies.

  • Current challenges: MVS performance can decline under noisy backgrounds or poor illumination, and complex production environments require algorithms adapted to different objectives and settings.These conditions can limit system robustness and reliability.
  • Current challenges: Food diversity makes it difficult to detect small colour changes and unusual shapes, while imaging technologies cannot readily support online smell detection.Smell remains an important reference feature for food safety and quality evaluation, but it is difficult to correlate appearance with smell.
  • Current challenges: Large image datasets increase the storage and computation requirements of MVS applications.
  • Future trends: Future development should improve image-processing efficiency and robustness through existing or new algorithms.Image processing is identified as the core of MVS and a prerequisite for future applications.
  • Future trends: Embedded vision systems are expected to expand MVS adoption because they offer compact structures, fast processing, and low cost.
  • Future trends: Future MVS research should further combine machine vision with multiple technologies, particularly as Internet of Things technology becomes prevalent.

6. Conclusion

The survey reviews machine vision systems, image processing, and machine learning applications in food processing. It concludes that MVS supports food safety inspection, monitoring, foreign-object detection, and more efficient work, while challenges remain and adoption is expected to increase.

  • The survey explains machine vision principles, image-processing steps, and recent food-processing applications.
  • MVS is widely used for food safety inspection, processing monitoring, foreign-object detection, and related tasks.
  • Machine learning can substantially increase the processing capacity of machine vision systems.
  • Although challenges remain, the survey identifies implementation of MVS in food processing as a general trend and expects its use to grow.
Loading 2103.16106v1…