Meta-learning for Few-Shot Insect Pest Detection in Rice Crop

Recent advancements in the field of Deep learning have helped in predicting and locating pests in agricultural field images accurately. A drawback of this approach is that it requires a large training dataset for each sample, which is not feasible. Since there is a wide variety of pests, collecting thousands of training images for each sample is impractical. To deal with this issue, a pest detection meta-learning technique based on Few-shot is proposed in this paper. In this work, pests from rice crops are considered for experiments. Two pest-image datasets: IP102 as a supported dataset to perform meta-learning and an image library for insects and pests known as the Indian Council of Agricultural Research-National Bureau of Agricultural Insect Resources (ICAR-NBAIR) are taken to perform Few-shot learning. In meta-learning phase, the proposed model is trained on a variety of pests, and hence the proposed system is capable of learning new categories of pests with very few training images.