Checking a meal with several adjacent foods
A prepared-meal camera has to distinguish neighbouring ingredients before it can decide whether something is missing. A September study tackles that problem with a relatively lightweight inspection system for Chinese-style lunch boxes. It combines colour features, a neural-network classifier and menu-specific thresholds to flag missing foods, insufficient portions and a simulated foreign object.
Under the experimental conditions, the authors report an overall image classification rate of 96.45%, a defective-image detection rate of 97.76% and a normal-image false alarm rate of 5.34%. These are different measures: detecting an abnormal meal is not the same as correctly identifying every ingredient or assigning every defect to the right category. The results describe the tested dataset, not a guaranteed production-line performance level.
Colour regions stand in for ingredient identity
The system first locates the meal container and filters the image to reduce bright reflections from oily surfaces. It then divides the image into fixed 6 × 6 regions and extracts statistical colour information in RGB, HSV and CIE Lab spaces. A deep neural network classifies these regions, and a filling step refines the resulting ingredient areas.
Those areas are compared with empirical thresholds for the expected menu. Rice, meat, vegetables, egg and pickled side dishes formed the food classes. The approach estimates whether the visible amount is sufficient; it does not directly weigh the meal or measure nutritional content. Changing a menu or its nominal serving size would require new data and recalibration rather than simply reusing the same thresholds.
The foreign-object test was deliberately narrow
For the contamination simulation, the researchers used a plastic cockroach model and placed pieces in different positions. An object counted as detected when more than half its manually annotated area was recognised. The paper identifies that rule as an engineering choice, not a regulatory or universal food-safety threshold.
Foreign-object region classification averaged 82.12% across five validation folds, even though meal-level decisions could perform better after information was combined across an image. Transparent plastic, hair, reflective metal, bone fragments and objects resembling the surrounding food were not comprehensively validated. The headline meal-classification percentage therefore must not be presented as a general contaminant detection rate.
Prototype feasibility still leaves a production test
Experiments varied lighting, container configurations and conveyor speed. Stronger lighting changes and motion blur reduced performance, especially for small or partly hidden ingredients. Egg and pickled-side-dish regions were harder to classify than larger, more visually distinct areas. The study also assessed individual defect categories separately, leaving meals with several simultaneous abnormalities unvalidated.
The authors call for larger independent datasets, repeated acquisition sessions and real production-line trials. A controlled comparison with modern segmentation systems is also still needed. For meal manufacturers, the work demonstrates a possible inspection architecture and makes its maintenance requirements visible: menu adaptation, threshold calibration and realistic contaminant testing are part of deployment, rather than details that a high laboratory accuracy score resolves automatically.
- Vision-Based Automated Inspection of Box Meals for Food Portion Defects and Foreign Object Detection. ↗
Multidisciplinary Digital Publishing Institute (MDPI) · CC-BY-4.0 ↗. Lin HD, Chen GM, Lin CH. Vision-Based Automated Inspection of Box Meals for Food Portion Defects and Foreign Object Detection. Sensors (Basel, Switzerland). 2026. DOI: 10.3390/s26175636. © 2026 by the authors. Adapted by FoodRadar.