基于视觉的苗期作物株间除草关键技术研究现状

Research status of key techniques of inter-plant weeding in seedling crops based on vision

基于视觉的苗期作物和杂草的图像分割技术逐渐成熟,通过视觉技术对苗期作物进行精准识别和定位,是实现株间除草的关键技术和难点。作物的精准识别首先需要利用颜色特征将图像中的作物、杂草和土壤背景进行分割;其次利用实际识别对象的位置特征,形状特征,纹理特征,光谱特征等构造新的特征向量,结合成熟的分类算法对作物和杂草进行特征分类识别。针对棉苗和大豆苗,主要提取位置特征、形状特征,多采用支持向量机为主分类算法;针对玉米,主要提取位置特征、纹理特征,多采用人工神经网络为主的分类算法;针对部分蔬菜苗,主要提取形状特征、光谱特征,多采用算法结合的优化算法,具体实现时需要根据离线样本学习的结果来平衡苗期作物的识别准确率与实时性。在目前的算法中,主要存在三方面的问题:作物特征提取效果易受到遮挡、光照等干扰;分类算法目前还不能得到非常令人满意的准确性和实时性;目前算法一般是针对某种时段的作物,不具有通用性。这些都是后续算法研究中需要进一步解决的问题。

Vision-based image segmentation technology of crops and weeds at seedling stage is gradually mature. Accurate recognition and positioning of seeding crops through visual technology is the key technology to achieve inter-plant weeding. Firstly, the color feature is used to segment the crops and weeds from soil background in the image. Secondly, the new feature vectors are constructed from the position feature, shape feature, texture feature and spectral feature of the actual recognized object, then selected classification algorithm with these features. Finally, find the most effective method to achieve target crop recognition. For cotton or soybean seedlings, the location and shape features are extracted and Support Vector Machine is usually adopted as the main classification algorithm. For corn, the location features and texture features are extracted and artificial neural network is usually adopted as the main classification algorithm;for some vegetable seedlings, shape features and spectral features are extracted and combine a variety of algorithms used compositely to recognize. According to the actual situation and offline sample learning results, the accuracy and real-time of seedling identification was balanced. There are three problems in current algorithms: The results of crop feature extraction are easily disturbed by occlusion, illumination and so on. The accuracy and real-time of classification algorithm are not good enough. It is not universal for the algorithm is aimed at crops for a certain period of time. These are the problems that need to be further discussed and solved in the follow-up algorithm research.