Deep learning-based fruit object detection and ripeness grading

Abstract

With the rapid development of smart agriculture, precision agriculture, and automated harvesting technologies, accurate fruit multi-object detection and ripeness grading have become critical research directions in agricultural intelligence. Traditional manual inspection methods suffer from low efficiency, strong subjectivity, large errors, and poor adaptability to complex environments, failing to meet the requirements of large-scale, intelligent, and fine-grained management in modern agriculture. To address challenges such as large variations in fruit target sizes, complex illumination conditions, and significant differences in ripeness levels across different environments, this paper proposes a fruit object detection and ripeness grading method based on an improved YOLOv10 model. Two independent detection models are trained separately: one for multi-category fruit object detection, the other for fruit ripeness stage grading. The two models can be deployed in series on embedded terminals to realize sequential fruit species recognition and ripeness evaluation. All raw images are randomly split into training, validation and test sets at a ratio of 7:2:1 before data augmentation, and augmentation operations are only applied to training samples to avoid data leakage. Taking YOLOv10-nano as the baseline framework, lightweight and feature enhancement improvements are made to the network architecture. The backbone network is replaced with MobileNetV3-Large, leveraging depthwise separable convolutions and inverted residual structures to achieve lightweight adjustment. A CBAM attention mechanism is introduced into the Neck module to adaptively enhance channel-wise and spatial features of multi-scale feature maps, thereby improving feature representation of fruit targets in complex backgrounds. A decoupled detection head is adopted to separate the classification task from the bounding box regression task, reducing feature conflicts between tasks and further improving the accuracy of fruit category recognition and ripeness grading. Experimental results demonstrate that the improved model maintains a lightweight architecture while achieving significantly enhanced detection performance. Specifically, the fruit object detection model achieves a mAP@0.5 of 0.88 and an inference speed of 78 FPS; the ripeness grading model achieves a mAP@0.5 of 0.87 and an inference speed of 74 FPS. The model parameters are reduced from 2.6M to 1.8M.