Sorting and counting of almond kernels on conveyor belt using computer vision and deep learning techniques

The classification and sorting of agricultural products, such as almonds, are critical processes in ensuring quality and meeting market demands. Sorting process is done in machines with computer vision technology. However, the output of these machines is never 100 %. The products coming out of these machines are finally sorted again on the conveyor belt. In this study, deep learning and computer vision techniques were used to perform the final sorting and counting on the conveyor belt. A self-curated dataset containing 1200 images divided into three distinct classes: whole almond kernels, damaged kernels, and broken shells was created to facilitate the study. We evaluated the performance of four CNN architectures: ResNet50, InceptionV3, VGG16, and EfficientNetB3 using both RGB and grayscale image datasets. Among these, EfficientNetB3 achieved the highest accuracy of 99.44 % with RGB images and 98.33 % with grayscale images. Field tests with new samples validated the model's robustness, achieving 97.14 % accuracy on RGB images and 95.71 % on grayscale images. These results demonstrate the potential of the proposed method to automate almond classification and sorting on the conveyor belt and calculate the operating accuracy of sortex machines with its counting feature.