Multi-dimensional field weed identification method based on generative adversarial learning

Abstract

The invention discloses a multi-dimensional field weed identification method based on generative adversarial learning. A semi-supervised learning strategy is adopted to train an overall network. The overall network is trained by inputting labeled data and unlabeled data. Tagged data is used as input of a discrimination network. training the network by using unlabeled data. A space probability graph generated by unsupervised data is used as calculation of semi-supervised loss. The method can effectively reflect the similarity between a prediction result and real data distribution, uses generative adversarial learning to find a confidence region of unlabeled data, obtains a final field weed recognition result through self-learning, uses a confidence graph to perform self-learning on a network, and uses a semi-supervised learning strategy to break through the limitation of insufficient labeled data. According to the invention, the segmentation network is used to replace a traditional network, input features of different sizes can be accepted, and the flexibility is high.

Description

technical field [0001] The invention belongs to the technical field of artificial intelligence, and relates to a method for identifying multi-dimensional field weeds based on generative confrontation learning. Background technique [0002] The existence of weeds poses a threat to the yield of crops in the field, and traditional indiscriminate chemical weeding methods have brought hidden dangers to the environment and food safety. Therefore, how to accurately distinguish weeds has become a key technology for precise weeding. [0003] The target recognition method based on computer vision has become the preferred method for target recognition due to its non-destructive, fast, and automatic characteristics. However, when this type of method is applied to the field of weeds in the field, it usually encounters two major problems: 1) Insufficient data labels , compared with traditional visual scenes, there are fewer existing datasets of labeled field weeds, and labeling them requi...

AI Technical Summary

Problems solved by technology

[0002] The existence of weeds poses a threat to the yield of crops in the field, and traditional indiscriminate chemical weeding methods have brought hidden dangers to the environment and food safety. Therefore, how to accurately distinguish weeds has become a key technology for precise weeding

[0003] The target recognition method based on computer vision has become the preferred method for target recognition due to its non-destructive, fast, and automatic characteristics. However, when this type of method is applied to the field of weeds in the field, it usually encounters two major problems: 1) Insufficient data labels , compared with traditional visual scenes, there are fewer existing datasets of field weeds labeled, and labeling them requires a lot of manpower, material and financial resources

Milioto, A., Lottes, P., & Stachniss, C..(2017). Real-time blob-wise sugar beets vs weeds classification for monitoring fields using convolutional neural networks using geometric transformation methods such as affine transformation, scale transformation and random cropping Image data is expanded. However, on the one hand, this type of method has limited performance improvement because the transformed data is too similar; result

2) The performance of the recognition model is poor. Compared with traditional visual scenes, weeds and crops have a higher similarity in color texture, Perez-Ortiz, M., Pen~a, J.M., Gutierrez, P.A., Torres-Sa'nchez ,J.,Herva′s-Mart′1nez,C.and Lo′pez-Granados,F.,2016.Selecting patterns and features for between-and within-crop-row weed mapping using uav-imagery.Expert Systems with Applications 47, pp.85–94. Weeds and crops are distinguished by statistical information of pixel vectors, but because the statistical information needs to be manually set, the performance varies on different data sets

Milioto, A., Lottes, P., & Stachniss, C..(2017). Real-time blob-wise sugar beets vs weeds classification for monitoring fields using convolutional neural networks. Field crops are identified by constructing CNN, but CNN requires pictures The data size is the same, the scope of application is too narrow

Images