Detection of early latent bruise on mangoes based on hyperspectral imaging combined with band operations and improved YOLOv5

Abstract

Early latent bruises in mangoes are difficult to detect before visible discoloration, yet they drive postharvest losses and quality downgrades. This study asks whether reflectance-calibrated hyperspectral imaging (HSI) combined with a lightweight, improved YOLOv5 can detect such non-visible lesions. Reflectance-normalized Vis/NIR HSI was acquired; ENVI-based PCA was used only to select characteristic wavelengths, one/two/three-band inputs were constructed via linear stretch, band-product, and RGB pseudo-color, and a YOLOv5n model enhanced with Ghost bottlenecks and Coordinate Attention was trained and evaluated using Ultralytics val. The resulting pipeline detected early latent bruises with high accuracy, achieving mAP_0.5 = 98.2% with balanced precision/recall and fast inference at 320×320 resolution. These findings are important for postharvest inspectors and automated sorting lines that require non-destructive, early decision-making to reduce waste and maintain supply-chain quality. Future research is motivated to collect pixel-level masks for segmentation metrics, assess cross-cultivar and condition generalization, and benchmark newer YOLO families under controlled protocols.

Abstract from the paper’s OpenAlex record; open copy licensed cc-by-nc-nd. Read at the source ↗

The text the model read, with the source record’s inline markup (italics, sub- and superscripts) rendered and its character entities decoded; quote offsets refer to the raw text.