Precision Almond Classification and Counting Using CNN Architectures and Blob Detection Algorithm

Abstract

The classification process, which is extensively employed in the food sector, is used to identify various product classes or to detect and sort solid, defective, and undesired objects in the harvested crop. In this study, blob detection algorithm and CNN architectures were used effectively to determine the number of “whole almonds”, “damaged almonds” and “almond shells” in almond samples. With 400 images in each class, a total of 1200 images were gathered. The final dataset was divided into training, validation, and testing sections at ratios of 70%, 15%, and 15%, respectively. The constructed dataset was used for training models such as VGG16, InceptionV3, ResNet50, and EfficientNetB3 architectures where EfficientNetB3 yielded the maximum accuracy of 99.44% for RGB dataset and 98.33% for grayscale dataset. To confirm the validation of the trained EfficientNetB3 architecture in the application, totally new 50 whole almonds, 10 damaged almonds and 10 shell images were acquired and the model was placed to the test. As a result of this experiment, the test accuracy was calculated as 97.14% for RGB image and 95.71% for grayscale image. As a result of the classification obtained from the RGB image, the model classified this data as 52 whole kernel almonds, 10 damaged numbers and 8 shells. These results show that the proposed method works in high accuracy with EfficientNetB3 model as a final application for both RGB and grayscale images.