基于改进YOLACT的果树叶墙区域实时检测方法

为了解决果园因农药过量使用导致的环境污染与农药浪费问题，提出了一种基于改进YOLACT的果树叶墙区域(Leaf wall area, LWA)实时检测方法，用于计算深度-彩色双目相机采集视频中的叶墙区域距离及密度，为果园农药智慧喷施作业中农药喷洒剂量与喷洒距离的实时调整提供依据。首先，使用ConvNeXt主干网络改进了YOLACT模型，并引入NAM通道注意力机制对模型进行了优化；其次，提出了基于深度学习的果树叶墙密度检测方法；最后，通过阈值法排除深度图像中的干扰信息，简化了果树叶墙平均距离计算方法的处理流程。实验结果表明，改进YOLACT模型分割的AP_all为91.6%,相较于原始模型上升3.0个百分点，与YOLACT++、Mask R-CNN和QueryInst模型相比分别高2.9、1.2、4.1个百分点；叶墙密度估计算法在叶墙顶部、中部和底部的均方根误差(Root mean square error, RMSE)分别为1.49%、0.82%、2.20%;叶墙区域实时检测方法的处理速度可达29.96 f/s。

Real-time Detection Method of Fruit Leaf Wall Area Based on Improved YOLACT

To reduce the environmental pollution and pesticide waste in orchards, a real-time method to detect the fruit tree leaf wall area (LWA) based on the improved YOLACT model was proposed to estimate average distance and density in the videos that captured by depth-color binocular camera, which can provide data for the real-time adjustment of pesticide spraying dose and spraying distance on intelligence pesticide spraying. Firstly, the YOLACT model was improved by using the ConvNeXt backbone network, and the NAM channel attention mechanism was introduced to optimize the model. Secondly, a leaf wall density estimation method based on deep learning was proposed. Finally, the average distance calculation method of LWA was proposed by excluding the interference information in the depth image through the threshold algorithm to simplify processing flow. The experimental results showed that the segmentation APall metrics of the improved YOLACT model was 91.6%, which was increased by 3.0 percentage points compared with that of the original model, and 2.9 percentage points, 1.2 percentage points and 4.1 percentage points compared with that of YOLACT++, Mask R-CNN, and QueryInst. The root mean square error (RMSE) of the leaf wall density estimation method was 1.49%, 0.82% and 2.20%. And the processing speed of the real time LWA detection method could reach 29.96f/s.