CSFN-YOLOv5s: A Rapid Apple Detection Algorithm in the Natural Environment

Abstract

In the natural environment, the rapid detection of apples is of great importance for picking robots. However, the complex growth conditions of apples, the occlusion of leaves and branches, and the distance can cause the problem of missed apple detection. To address this problem, an algorithm called CSFN-YOLOv5s for accurate and efficient apple detection in complex natural environments was developed. Firstly, the Context Augmentation Module - Spatial Pyramid Pooling with Feature Concatenation and Spatial Channel wise Pooling （CAM-SPPFCSPC ）framework is constructed to introduce additional background and context information, enhance the use of context information of the model, and help the model better understand the image information, so as to improve the robustness and generalization ability of the model. Secondly, with the application of four detection layers to obtain finer granular feature expression and a smaller receptive field, improve the accuracy of small target detection by finely capturing its detailed information. Thirdly, the Normalized Wasserstein Distance (NWD) was used to improve the sensitivity of IoU to the position deviation of small objects. Finally, a large number of experimental results show that CSFN-YOLOv5s has certain advantages for rapid identification of apples in natural and complex environments.