Abstract:

To address the issues prevalent in existing apple instance segmentation algorithms, including variable object scales, susceptibility to background noise which degrades segmentation accuracy, and excessively slow inference speed, an improved SOLOv2-based apple instance segmentation algorithm, termed AE-SOLOv2, is proposed in this paper. First, the StarBlock structure is adopted to optimize the feature extraction network, thereby improving inference efficiency. Subsequently, an enhanced bidirectional feature pyramid network (E-BiFPN) is designed, where additional cross-layer connections facilitate adequate fusion of high-level semantic information and low-level geometric detail information, mitigating the accuracy degradation caused by inconsistent scales of segmentation targets. Furthermore, an attentive feature fusion module is introduced to integrate multi-level and multi-branch feature information. Combined with an adaptive grouping strategy and collaborative processing of global and local features, this module significantly enhances the anti-interference capability of the algorithm against background noise. Experimental results demonstrate that the improved algorithm effectively enhances the instance segmentation precision for apples. On the validation dataset, the mean average precision at an intersection-over-union threshold of 0.5 (mAP₅₀) reaches 91.2%, and the inference time per image is 12.3 ms, which satisfies the real-time requirements for apple instance segmentation. In addition, picking experiments verify that AE-SOLOv2 exhibits superior overall performance even in apple-picking robot scenarios with constrained computational resources, providing a novel solution for the development of agricultural intelligence.

Key words: SOLOv2 network, instance segmentation, picking robot, attention mechanism