Abstract Object detection in natural environments is a critical challenge for agricultural automation, particularly for small and occluded targets such as apples among foliage. We propose an innovative detection framework CSFN-YOLOv5s. Firstly, the context augmentation module-spatial pyramid pooling faster cross stage partial channel structure is constructed to introduce additional background and context information, so that the model can understand the image data more deeply and improve its robustness and generalization ability. Secondly, four detection layers are applied to obtain finer-grained feature expression and smaller receptive field, which improves the precision of small target detection by finely capturing the details. Thirdly, normalized wasserstein distance is used to mitigate the sensitivity for small object localization errors, showing a significant improvement. The experimental results and analysis show that the mAP of the final model reaches 98.5%, which is significantly better than other mainstream target detection models. Especially for the small target detection task in the natural complex environment, the proposed CSFN-YOLOv5s model shows higher detection precision and verifies its target detection effect.
Read full abstract