TRD-Net: an efficient tomato ripeness detection network based on improved YOLO v8 for selective harvesting

Fruit recognition and ripeness detection are crucial steps in selective harvesting. To better address the difficulties of fruit recognition and ripeness detection techniques in complex facility environments, a novel lightweight tomato ripeness detection network model based on an improved YOLO v8s is proposed (called TRD-Net). Here, a tomato dataset including 3,330 images from real scenarios was constructed, and an accurate lightweight tomato ripeness detection model trained on the captured images was developed. The TRD-Net model achieves efficient detection of tomatoes affected by overlapping occlusions, lighting variations, and capture angles, offering swifter detection speeds and lower computational demands. Specifically, the feature extraction module of YOLO v8s was refactored by employing spatial and channel reconstruction convolution (SCRConv) and adding the SimAM attention mechanism. The CIoU loss function was replaced by the MPDIoU loss function. The performance of the novel TRD-Net was comprehensively investigated. The proposed TRD-Net achieved an mAP@0.5 of 0.9581 with an improvement of 4.32 percentage points, and the model size decreased from 22.5 M to 17.6 M with an inference time of 8.7 ms per image. The number of model parameters and floating-point operations per second (FLOPs) decreased by 19.69% and 22.03%, respectively. Compared with state-of-the-art models, the proposed TRD-Net is notably promising for real-time tomato recognition and maturity detection. The study contributes to the establishment of a machine vision sensing system for a selective harvesting robot in a complex gardening environment.