Apple estimation and recognition in complex scenes using YOLO v8

Abstract

Apple-picking robots are designed to accurately identify ripe apples, efficiently harvest them, and offer adaptable solutions applicable to various fruits. However, current methods show limited detection accuracy in complex environments due to challenges such as shading in natural settings, variations in ambient lighting, and diversity in fruit color, shape, and size. This study integrates Hue–Saturation–Value (HSV) color space transformation with YOLO v8 models to identify the quantity and location of apples, while also assessing their ripeness and weight. First, image data of apples in natural environments were collected and preprocessed using grayscale conversion and Gaussian filtering for denoising. Next, edge detection was performed using the Canny and Sobel algorithms, and apple counting was achieved via the YOLO v8 model. Ripeness was evaluated based on HSV and Red–Green–Blue (RGB) values, while weight estimation was conducted by constructing 3D models of the fruits. Finally, through color feature extraction and YOLO v8, apples were precisely identified among various fruits. Experimental results show that the YOLO v8 model achieved an average precision of 95% and an F1-score of 93.45%. Compared to existing algorithms such as Visual Geometry Group (VGG) and Residual Network (ResNet), YOLO v8 improved Mean Average Precision (mAP) by 0.14% and 8%, respectively. These advancements provide strong technical support for apple-picking robots, enabling faster and more accurate operations, improving fruit quality, and meeting the practical demands of agricultural production.