Detecting defects on cheese using hyperspectral image analysis

Abstract: Defects such as mold and bacterial stains can appear on cheese. Manually detecting defects is a time-consuming, cost-ineffective, and ergonomically unsatisfactory process for a dairy because the quality technicians must inspect each cheese before packaging. Instead, dairies would prefer an automatic detection system, but it is unclear whether reliable options are available. One potential approach is hyperspectral image analysis, which can interpret and classify chemical information from a sample. We collected hyperspectral images from a dairy using a short-wave infrared (SWIR) camera and compared three prediction models: a PLS-discriminant analysis with the software Breeze from the analysis company Prediktera and two classifiers based on a convolutional neural network and a support vector machine, both coded in Python. We found mold and lactobacilli stains using all methods, but dirt was more challenging to detect. Also, the methods had issues with false positives of lactobacilli stains. To improve accuracy, we recommend collecting more data, especially samples with lactobacilli stains and dirt, and using more non-defect cheese for validation. To find smaller defects, we propose that future work should test a visible and near-infrared (VNIR) camera with higher resolution. Though PLS-discriminant analysis did not achieve the highest accuracy, it was not far off and had the most time-effective predictions. Since Breeze already integrates PLS-discriminant analysis, it should remain in focus, but for higher accuracy Prediktera should continue to explore other methods such as neural networks.