Aphids pose a significant threat to crop yield and quality, leading to economic concerns in agricultural production. Traditional pest monitoring methods are time-consuming, prompting the need for efficient visual image-based techniques. In outdoor data collection, obtaining large amounts of data for deep learning can be challenging. Therefore, we opted for a Few-shot learning approach. Our study focuses on Few-shot learning to quickly and adaptively identify aphid species with minimal samples. Utilizing prototypical networks, this study investigates two dataset compositions: a 3-class set (aphid, insects, and background) and a 9-class set (2 aphid classes, 6 insect classes, and background). We apply 3 different CNNs as the backbone of Prototypical Networks to both datasets creating 6 scenarios. For each scenario, we evaluate and compare the performance of CNNs. This involves assessing 6 different distance methods, two of which are newly introduced compared to prior research. The goal is to examine the effectiveness of backbones and distance methods in diverse experimental settings and find a reliable way to detect aphids using a few samples. The experiment with Resnet50, 3-class dataset, and cosine similarity method, as one of our scientific contributions, showed impressive recall values of 98% for aphid classification and 99% for overall insect classification. Notably, Euclidean distance and Mahalanobis divergence proved reliable, with Euclidean being more computationally efficient. Larger CNNs yielded superior results, but their resource demands should be considered. The 9-class dataset outperformed the 3-class dataset, highlighting the importance of data richness. Our findings affirm the reliability of prototypical networks as a dependable few-shot learning method, achieving satisfactory levels of aphid detection.
Link to the paper of this project.
