Detection of bruised apples using structured light stripe combination image and stem/calyx feature enhancement strategy coupled with deep learning models

Abstract

This study presented a novel approach that integrated visible structured-illumination reflectance imaging (SIRI) with deep learning techniques to concurrently identify the stem, calyx, and bruise in apple. Structured light images of apple samples were acquired at five frequencies (0.15, 0.20, 0.25, 0.30, and 0.50 cycles mm⁻¹) at four time points (0, 6, 12, and 24 ​h) using a developed SIRI. A three-step phase-shifting method was then applied to demodulate the images to obtain the direct component (DC), alternating component (AC), and the ratio (RT) images. Independent stripe images were extracted and skeletonized, and superimposed onto the original AC and RT images to generate a composite image with enhanced stem/calyx features. Three deep learning models (Faster R–CNN, YOLO-v5s, and YOLO-v8n) were used to recognize apple stem/calyx and bruise regions. The study showed that the composite image with an optimal frequency of 0.30 cycles mm⁻¹ can improve recognition accuracy. Among the three models, the YOLO-v8n achieved the highest classification accuracy (99.12%) for detecting bruised apples.