Lightweight and precise potato slice quality evaluation using a CNN–LightGBM hybrid approach

Potatoes are among the major vegetables grown worldwide due to their vast usage in the Fast-Moving Consumer Goods (FMCG) Industry. During industrial processing, potato slices used in prepared food products can develop defects that compromise the final product’s quality. The manual inspection method of potato processing takes a very long time and is more vulnerable to human error, leading to large amounts of wasted food. To overcome these challenges, the research proposes an advanced defect detection system using an ensemble of deep learning models, namely DenseNet121 and EfficientNetB3, along with LightGBM as a metamodel. DenseNet121 is employed for discriminative feature extraction and defect classification, while EfficientNetB3 further enhances performance through robust feature learning and accurate classification. LightGBM, known for its speed and high efficiency in handling large-scale data with lower memory usage, is particularly effective at capturing complex non-linear relationships and ranking features. The model is trained and evaluated on a custom-built dataset containing healthy and defective potato slices, including both isolated slices and grouped formations. The proposed method aims to decrease food waste, improve quality control, and achieve higher efficiency in production industries. The proposed method is validated with an accuracy of 99.83%, precision of 99.85%, recall of 99.78%, and F1-score of 99.81% for reliable defect detection in potato slices. The research contributes to reducing food waste, enhancing production efficiency, and supporting consistent quality assurance in the food processing industry.