基于YOLO 11n-LDA的苹果成熟度识别与低损采摘方法

Apple Ripeness Detection and Low-damage Harvesting Based on YOLO 11n-LDA

摘要:

针对不同光照下苹果成熟度识别困难,不同成熟度苹果使用同一采摘模式损伤率较高的问题,提出了一种改进 YOLO 11n 不同光照下苹果成熟度识别模型YOLO 11n-LDA,并与柔性夹爪结合实现不同成熟度苹果低损采摘的方法。首先,在YOLO 11n网络中的SPPF模块中加入具有大型可分离内核注意力(Large separable kernel attention, LASK)模块,增大模型感受野;然后在C3k2模块中加入动态蛇形卷积(Dynamic snake convolution,DySnakeConv),提高模型多尺度特征感知能力;其次将Neck网络替换为渐进式特征金字塔(Asymptotic feature pyramid network,AFPN)网络,可以减少语义间隙,提高模型特征融合的稳定性。最后使用ANSYS仿真分析出不同成熟度苹果的最大安全采摘力,融合到柔性夹爪中实现低损采摘。在PApple_RGB-D-Size数据集中苹果分为食用成熟、采摘成熟、未成熟3类, 改进后模型整体的平均精度均值(mAP@ 0. 5)为87. 7% ;食用成熟、采摘成熟和未成熟的平均精度均值(AP@ 0. 5)分别为86. 6% 、89. 0% 和87. 5% ;整体的mAP@ 0. 5 和3 类成熟度的AP@ 0. 5 相较于YOLO 11n 分别提高3. 9、8. 0、1. 6、 2. 1 个百分点。在Orchard apple maturity 数据集中改进模型的mAP@ 0. 5 为95. 5% ,相较于YOLO 11n提高0. 9 个百 分点。采摘实验中检测率、采摘成功率和损伤率分别为99. 0% 、84. 0% 和4. 8% 。

Abstract:

Aiming to address the issue of high damage rates caused by a single harvesting mode for apples of varying maturities, an apple maturity identification model, YOLO 11n-LDA, was proposed based on YOLO 11n, combined with a flexible gripper to achieve low-damage harvesting for apples at different maturity stages. Firstly, the large separable kernel attention ( LSKA) module was integrated into the SPPF module of the YOLO 11n network to expand the model??s receptive field. Subsequently, dynamic snake convolution (DySnakeConv) was added to the C3k2 module to enhance the model??s perception of multi-scale features. Furthermore, the asymptotic feature pyramid network (AFPN) replaced the neck network to bridge the semantic gap and improve feature fusion stability. Finally, the maximum safe picking force for apples of different maturities was determined via ANSYS simulation, and a comprehensive harvesting test platform, comprising a collaborative robotic arm, a flexible fin gripper, a camera, and an industrial control computer was constructed for experimental testing. Experimental results on the PApple_RGB-D-Size dataset showed that, with maturity classified into eating-ripe, harvest-ripe, and unripe, the improved model achieved an overall mean average precision ( mAP @ 0. 5) of 87. 7% . The AP @ 0. 5 for eating-ripe, harvest-ripe, and unripe apples were 86. 6% , 89. 0% , and 87. 5% , respectively; the overall mAP@ 0. 5 and the AP@ 0. 5 for the three maturity categories was increased by 3. 9, 8. 0, 1. 6, and 2. 1 percentage points compared with that of YOLO 11n. In the orchard apple maturity dataset, the improved model achieved an mAP@ 0. 5 of 95. 5% , representing an increase of 0. 9 percentage points compared with that of YOLO 11n. In harvesting experiments, the detection rate, harvesting success rate, and damage rate were 99. 0% , 84. 0% , and 4. 8% , respectively. The enhancement of the vision model improved detection accuracy, while the visual-tactile fusion method provided a reference scheme for low-damage harvesting.