Strawberry ripeness detection based on improved lightweight network in cluttered field environment

To address inaccurate strawberry recognition caused by cluttered field environments such as varying illumination, occlusion and uneven distribution, an improved lightweight model YOLOv7-SSC for strawberry ripeness detection was proposed. First, the backbone network of YOLOv7 is replaced with ShuffleNetV2, a lightweight feature extraction network, to significantly reduce the number of model parameters. Second, the lightweight network Slim-neck is used as the neck structure to reduce model complexity while preserving high precision. Finally, the Content Perception Feature Recombination (CARAFE) upsampling is used to enlarge receptive field in the feature fusion network and fully leverage semantic information. Moreover, the pictures of the strawberry dataset with three common conditions (unripe strawberry, near ripe strawberry and ripe strawberry) were collected in real picking environment. The experimental results show that compared to the original YOLOv7 model, the improved model parameters are reduced by 69.0 %, the floating point number is decreased by 79.4 %, and the accuracy rate reaches 99.6 %. These results demonstrate that YOLOv7-SSC model can achieve fast recognition of strawberry maturity while maintaining high precision, making it more suitable for small target detection in complex field compared with other algorithms.