SFYOLO: a Lightweight and Effective Network Based on Space-Friendly Aggregation Perception for Pear Detection

EasyChair Preprint 8950

Abstract

It is always challenging for efficiently conducting accurate detection of small and occluded pears in modern orchards. In the past few years, the aforementioned detection tasks remained unsolved though lots of researchers attempted to optimize the adaption of background noise and viewpoints, particularly compliant models suitable for simultaneously detecting small and occluded pears with low computational cost and memory usage. In this paper, we proposed a lightweight and effective object detection network called as SFYOLO based on space-friendly aggregation perception. Specifically,a novel space-friendly attention mechanism was proposed for implementing the aggregate perception of spatial domain and channel domain. Afterwards, an improved space-friendly transformer encoder was put forward for enhancing the ability of information exchange between channels. Finally, the decoupled anchor-free detectors were used as the head to improve the adaptability of the network. The mean Average Precision (mAP) for in-field pears was 93.12\% in SFYOLO, which was increased by 2.03\% compared with original YOLOv5s. Additional experiments and comparison were carried out considering newly proposed YOLOv6 and YOLOv7 that aimed at optimizing the detection accuracy and speed. Results verified that small and occluded pears could be detected fast and accurately by the competitive SFYOLO network under various viewpoints for further orchard yield estimation and development of pear picking system.

Keyphrases: Aggregate perception, Transformer encoder, YOLOv5s, object detection, visual attention mechanism

@booklet{EasyChair:8950,
  author    = {Yipu Li and Yuan Rao and Xiu Jin and Zhaohui Jiang and Lu Liu and Yuwei Wang},
  title     = {SFYOLO: a Lightweight and Effective Network Based on Space-Friendly Aggregation Perception for Pear Detection},
  howpublished = {EasyChair Preprint 8950},
  year      = {EasyChair, 2022}}