Binary classification of defects in multiple coffee beans using lightweight convolutional neural networks for embedded systems

Abstract

Coffee and its derivatives are part of the world's gastronomy and are also one of the main export products in Imbabura, Ecuador. Coffee producers and marketers must ensure the quality of their coffee beans, as a single seed can present several morphological defects. However, the selection of quality beans is a manual, tedious, and time-consuming process, prone to misclassification due to human limitations such as fatigue and varying subjective classification criteria. This work aims to detect and classify multiple coffee beans automatically present in an image (both good and bad) using modern convolutional neural networks, specifically YOLOv11 (Nano and Small variants), which are selected for their computational efficiency and good performance in real-time detection tasks. A custom dataset with a total of 4,090 RGB images with multiple coffee beans randomly located within each image was collected and manually labeled into two classes (good and bad) using the CVAT tool. Image preprocessing included background variation (white, gray, and black) to improve dataset variability and model robustness. Models were trained on the Kaggle platform with an NVIDIA Tesla P100 GPU (16 GB VRAM), using the PyTorch framework and the Ultralytics library. Results showed an accuracy of 0.880 for YOLO v11-Nano and 0.871 for YOLO v11-Small, along with inference times of 7.89 ms (126.7 fps) and 10.08 ms (99.20 fps), respectively. These results are comparable to those reported in the literature, but now considering the challenge of superimposing multiple coffee beans randomly located in each image with a variable background. These results suggest the potential for applying the model to embedded systems in practical settings for small and medium-sized producers and associations, thereby contributing to technological innovation in the Ecuadorian coffee sector.