A lightweight CBBox-KiwiYOLO network for kiwifruit detection in open orchard conditions

Accurate target detection is a prerequisite for the operational effectiveness of kiwifruit harvesting robots. However, factors such as densely distributed fruits, inconsistent scales within view field, and complex lighting conditions pose significant challenges to this critical process. This paper proposes a CBBox-KiwiYOLO network. A CBBox was designed to match the spherical shape of fruits. To address the small size and dense distribution of kiwifruits, an SKAttention module was introduced, using SKNet to adaptively adjust kernel sizes for multi-scale feature extraction, with SENet calculating attention weights to enhance small target detection. CSWin transformer was integrated to improve long-range dependency perception. Additionally, a LEM module was designed to enhance adaptability under complex lighting conditions. The results show that CBBox-KiwiYOLO improves mAP50, Precision, and Recall by 3.86 %, 3.09 %, and 4.59 %, respectively, compared to the original YOLOv8. Tests from extreme lighting conditions validate the better performance of the proposed network. With a model size of only 6.3 M, an inference time of 19.6 ms per frame, and a video inference rate of 51 FPS, this network delivered competitive overall performance compared to more complex and time-consuming ones. These findings provide theoretical and technical support for the automated harvesting of kiwifruit.