Cognitive computing for classification of six weed species in tomato and maize crops

Agriculture has used technology as a tool to reinvent itself over time. Cognitive computing is used in numerous artificial intelligence applications, including expert systems, robotics, virtual reality, and machine learning. Supervised deep learning models require a large amount of labeled data for training, representing a common challenge that affects current studies of weed detection and classification using images acquired from a drone. The data augmentation technique provides a practical solution to alleviate insufficient datasets. In this study, we investigate the capability of a cognitive computational system applied to a real-world scenario for data generation, we use the pre-trained Generative Face Priority (GFP) model, which is employed in blind face restoration by means of spatial feature transformation layers, and this model is based on Generative Adversarial Neural Networks (GAN). Data augmentation was applied for weed species found in the early stages of maize and tomato crops; monocots (Cyperus rotundus, Lolium) and dicots (Atriplex patula, Convolvulus arvensis, Salsola kali and Solanum nigrum). The RGB images were acquired from a drone at an altitude of 12 m. The real and augmented datasets were systematically evaluated with those generated by GFP-GAN, for different scaling factors in the images that constitute the input to the classifiers based on the convolutional neural networks VGG16 and ResNet152. Our results show that the technique used improves the generalization capability of the evaluated classifiers, so the size of the dataset directly influences the performance of the models, noticing an approximate increase of more than 5% in the model accuracy metric; furthermore, the GFP-GAN architecture restores with realistic and faithful details by increasing the spatial resolution of the image. The proposed approach demonstrates the scope of the cognitive system built in a domain different from the target domain.