Abstract
Fruit detection and recognition has an important impact on fruit and vegetable harvesting, yield prediction and growth information monitoring in the automation process of modern agriculture, and the actual complex environment of orchards poses some challenges for accurate fruit detection. In order to achieve accurate detection of green fruits in complex orchard environments, this paper proposes an accurate object detection method for green fruits based on optimized YOLOX_m. First, the model extracts features from the input image using the CSPDarkNet backbone network to obtain three effective feature layers at different scales. Then, these effective feature layers are fed into the feature fusion pyramid network for enhanced feature extraction, which combines feature information from different scales, and in this process, the Atrous spatial pyramid pooling (ASPP) module is used to increase the receptive field and enhance the network's ability to obtain multi-scale contextual information. Finally, the fused features are fed into the head prediction network for classification prediction and regression prediction. In addition, Varifocal loss is used to mitigate the negative impact of unbalanced distribution of positive and negative samples to obtain higher precision. The experimental results show that the model in this paper has improved on both apple and persimmon datasets, with the average precision (AP) reaching 64.3% and 74.7%, respectively. Compared with other models commonly used for detection, the model approach in this study has a higher average precision and has improved in other performance metrics, which can provide a reference for the detection of other fruits and vegetables.