Non-destructive Detection and Classification of Multiple Lactuca sativa Varieties with Leaf Occlusions Using Deep Learning

Lettuce (Lactuca sativa) can lose its freshness quality during post-harvest due to degradation caused by chemical reactions. Preserving its freshness requires proper storage conditions which put a further financial burden on traders. However, such a burden can be reduced if the lettuce does not require prolonged storage time but having a non-destructive monitoring system is a favorable option. Determining the freshness quality of the lettuce through computer vision is the optimal choice for the monitoring system; but detecting the lettuce and classifying it with occluded leaves according to its variety is the initial step and also the focus of this paper. The object detection models trained in this study are Roboflow 3.0, YOLOv8, and TensorFlow which are used to detect the lettuce and classify whether it is green ice, iceberg, or romaine. Before training, the dataset was annotated and was pre-processed and augmented with resizing, auto-orienting, flipping, rotation, cropping, blurring, adding noise and cutout, and changing hue, saturation, and brightness. The result showed YOLOv8 having the least box loss and training loss, but no one model outperformed other models in all precision, recall, and mean average precision (mAP). Thus, by having a composite metric, it was determined that TensorFlow was the best model with a 93.40% mAP. While the results can be further improved through training with a wider dataset, it can already be used as a foundation for determining the freshness quality of the lettuce through computer vision.