基于轻量卷积结合特征信息融合的玉米幼苗与杂草识别

Recognition of Maize Seedling and Weed Based on Light Weight Convolution and Feature Fusion

针对自然环境下作物与杂草识别精度低、实时性和鲁棒性差等问题,以幼苗期玉米及其伴生杂草为研究对象,提出一种基于轻量卷积神经网络结合特征层信息融合机制的改进单步多框检测器(Single shot multibox detector,SSD)模型。首先,采用深度可分离卷积结合压缩与激励网络(Squeeze-and-excitation networks,SENet)模块构建轻量特征提取单元,在此基础上通过密集化连接构成轻量化前置基础网络,替代标准SSD模型中的VGG16网络,以提高图像特征提取速度;然后,基于不同分类特征层融合机制,将扩展网络中深层语义信息与浅层细节信息进行融合,融合后的特征图具有足够的分辨率和更强的语义信息,可以提高对小尺寸作物与杂草的检测准确率。试验结果表明,本文提出的深度学习检测模型对自然环境下玉米及其伴生杂草的平均精度均值为88.27%、检测速度为32.26 f/s、参数量为8.82×10^6,与标准SSD模型相比,精度提高了2.66个百分点,检测速度提高了33.86%,参数量降低了66.21%,同时对小尺寸目标以及作物与杂草叶片交叠情况的识别具有良好的鲁棒性与泛化能力。本文方法可为农业自动化精准除草提供技术支持。

The drawbacks of traditional crop and weed identification algorithms include low accuracy,poor real-time and weak robustness,resulting in weeding operations inefficient in the natural environment.In order to solve these problems,corn and its associated weed were taken as research object,and an improved single shot multibox detector(SSD)model was proposed.Firstly,a light weight feature extraction unit was constructed through the use of depth separable convolution and squeeze-and-excitation networks(SENet)module.On this basis,a light weight basic network formed with dense connection was adopted to replace the VGG16 network of the standard SSD model,so as to improve the speed of image feature extraction.Based on the mechanisms of different classification feature layer fusion,the deep semantic information in extra feature layers was fused with shallow detail information.The fused feature map would have enough resolution and strong semantic information,which can improve the detection accuracy of small-scale crops and weeds.Experimental results showed that the mean average precision and recognition speed of the proposed model were 88.27%and 32.26 f/s,respectively,and the parameters size was 8.82×106.Compared with that of standard SSD model,the identification accuracy and speed of this model were increased by 2.66 percentage points and 33.86%,respectively,and the parameters were decreased by 66.21%.In addition,the improved SSD model performed good robustness ability under the condition of small-scale targets and overlapping of crop and weed leaves.The proposed method could identify crop and weed accurately and rapidly,which provided a technical support for agricultural automatic precision weeding.