Wei Jinfan, Sun Yu, Luo Lan, Ni Lingyun, Chen Mengchao, You Minghui, Mu Ye, Gong He
College of Information Technology, Jilin Agricultural University, Changchun, China.
Jilin Province Intelligent Environmental Engineering Research Center, Changchun, China.
Front Plant Sci. 2025 May 2;16:1503256. doi: 10.3389/fpls.2025.1503256. eCollection 2025.
In order to meet the urgent need of fruit contour information for robot precision picking in complex field environments (such as light changes, occlusion and fruit overlap, etc.), this paper proposes an improved YOLOv8s-seg method for tomato instance segmentation, named ACP-Tomato-Seg. The method proposes two innovative modules: the Adaptive and Oriented Feature Refinement module (AOFRM) and the Custom Multi-scale Pooling module (CMPRD) with Residuals and Depth. By deformable convolution and multi-directional asymmetric convolution, the AOFRM module adaptively extracts the shape and direction features of tomatoes to solve the problems of occlusion and overlap. The CMPRD module uses the pooling kernels of self-defined size to extract multi-scale features, which enhances the model's ability to distinguish tomatoes of different sizes and maturity levels. In addition, this paper also introduces a partial self-attention module (PSA), which combines channel attention and spatial attention mechanism to capture global context information, improve the model's ability to focus on the target region and extract details. To verify the validity of the method, a dataset of 1061 images of large and small tomatoes was constructed, covering six ripened categories of large and small tomatoes. The experimental results show that compared with the original YOLOv8s-seg model, the performance of ACP-TomatoSeg model is significantly improved. In the bounding box task, mAP50 and MAP50-95 are improved by 5.6% and 8.3%, respectively, In the mask task, mAP50 and MAP50-95 increased by 5.8% and 8.5%, respectively. Furthermore, additional validation on the public strawberry instance segmentation dataset (StrawDI_Db1) indicates that ACP-Tomato-Seg not only exhibits superior performance but also significantly outperforms existing comparative methods in key metrics. This validates its commendable generalization ability and robustness. The method showcases its superiority in tomato maturity detection and fruit segmentation, thus providing an effective approach to achieving precise picking.
为满足复杂田间环境（如光照变化、遮挡和果实重叠等）下机器人精准采摘对果实轮廓信息的迫切需求，本文提出了一种用于番茄实例分割的改进YOLOv8s-seg方法，名为ACP-Tomato-Seg。该方法提出了两个创新模块：自适应定向特征细化模块（AOFRM）和带残差与深度的自定义多尺度池化模块（CMPRD）。通过可变形卷积和多方向不对称卷积，AOFRM模块自适应提取番茄的形状和方向特征，以解决遮挡和重叠问题。CMPRD模块使用自定义大小的池化核提取多尺度特征，增强了模型区分不同大小和成熟度水平番茄的能力。此外，本文还引入了部分自注意力模块（PSA），它结合了通道注意力和空间注意力机制来捕捉全局上下文信息，提高模型聚焦目标区域和提取细节的能力。为验证该方法的有效性，构建了一个包含1061张大、小番茄图像的数据集，涵盖大、小番茄的六个成熟类别。实验结果表明，与原始YOLOv8s-seg模型相比，ACP-TomatoSeg模型的性能有显著提升。在边界框任务中，mAP50和MAP50-95分别提高了5.6%和8.3%；在掩码任务中mAP50和MAP50-95分别提高了5.8%和8.5%。此外，在公共草莓实例分割数据集（StrawDI_Db1）上的额外验证表明，ACP-Tomato-Seg不仅表现出卓越性能，而且在关键指标上显著优于现有对比方法。这验证了其值得称赞的泛化能力和鲁棒性。该方法在番茄成熟度检测和果实分割方面展现出优越性，从而为实现精准采摘提供了一种有效方法。
