Particle Size Detection of Palm Kernel Cake from Sieving Based on Images Using Convolutional Neural Network

Abstract

Palm kernel cake (PKC), a by-product of the palm oil industry, is widely used in animal feed due to its economic value. Its utilization reduces the reliance on costly conventional feed ingredients, reducing production expenses and improving livestock efficiency. However, contamination with palm kernel shells remains a key challenge, as it reduces quality and nutritional value. Identifying PKC particle sizes and addressing inconsistencies caused by contamination is complex, requiring advanced computational solutions. This study focuses on classifying the PKC particle sizes -fine, medium and coarse - using image processing combined with machine learning. A sieve shaker is applied to separate particles by size distribution, and a classification model is developed with Convolutional Neural Networks (CNN) under a transfer learning framework, which is effective for limited datasets. Six CNN architectures, MobileNet, Xception, InceptionV3, ResNet-152, VGG16, and NasNetMobile, are tested in four-layer configurations to identify the optimal setup. The results show that the proposed approach can classify PKC particle sizes with high accuracy. Among the models tested, MobileNet provides the best performance, achieving 0.99 accuracy and 0.98 F1 score in the second variation experiment. These findings present a practical and cost-effective method for assessing the quality of PKC, supporting scalable applications in feed production. This approach not only improves the accuracy of the evaluation, but also contributes to efficiency and sustainability in the livestock industry.