Multistage AI Crop Inspection and Segmentation Framework (MACISF)

The Multistage AI Crop Inspection and Segmentation Framework (MACISF) is a cutting-edge automated pipeline designed to revolutionize fruit quality assessment in modern agriculture. This innovative system bypasses the limitations of conventional manual inspection, which is both time-consuming and subjective, by employing a sophisticated deep learning approach.

MACISF's architecture is divided into three distinct stages, each tailored to a specific aspect of fruit quality evaluation. The first stage employs an EfficientNetV2-L backbone, which has been fine-tuned through a three-stage progressive strategy to accurately classify crop types. This sophisticated approach ensures that MACISF can differentiate between various fruit species with exceptional precision, setting it apart from other models like InceptionV3, ResNet-50, Xception, EfficientNetB3, DenseNet-121, and MobileNetV2.

In the second stage, MACISF utilizes seven species-specific EfficientNetV2-S classifiers, each meticulously trained to distinguish between healthy and defective fruit samples. This highly targeted strategy guarantees a mean health state accuracy of 99.39% and a mean defective-class recall of 99.51%, enabling MACISF to deliver comprehensive insights into the overall quality of the fruit crop.

The third and final stage of MACISF incorporates a YOLOv8m module, which specializes in defect localization. By pinpointing the exact location of defects within the fruit, this module enables farmers to address specific issues and optimize their post-harvest management strategies.

MACISF's capabilities extend far beyond simple classification, as it also provides a per-unit economic and nutritional evaluation, allowing farmers to make informed decisions about their crop's economic value and nutritional content. This multifaceted approach not only enhances yield optimization but also contributes to improved food security for communities worldwide.

To validate its efficacy, MACISF was trained and evaluated on a composite dataset comprising seven fruit species. The results speak for themselves, with the framework achieving an impressive 99.23% crop type accuracy, as measured by both macro F1-score (99.23%) and macro AUC-ROC (0.993). Additionally, MACISF's projected grading mAP@0.5 stands at an impressive 92.3%, underscoring its ability to deliver highly accurate and reliable results.