Ai-Driven Detection of Tiny Pests in Foliage: Integrating Image Processing and Deep Learning

Abstract

We present a novel computer vision method for detecting insect pests on plant or tree leaves under real-world conditions combining deep learning and classical image processing techniques. Detecting small insects that are sparsely distributed or camouflaged is challenging, as current state-of-the-art object detection methods, primarily designed for larger objects, often overlook them. Our approach to addressing this issue is twofold. To detect insects, we utilize a deep learning model that analyzes suspicious leaves for anomalies—a task well-suited to deep learning. However, deep learning struggles with tiny objects in complex backgrounds. To address this, we use conventional image processing to pre-identify potentially infested foliage, guiding the model to relevant areas and mitigating its limitations. This combined approach proves effective and competitive with other methods across various datasets and real-world scenarios, while also not requiring extensive training data. A thorough analysis is conducted to interpret the deep learning model's predictions, boosting confidence in its effectiveness for agricultural pest management. The proposed system ultimately enhances automated pest identification and expands its applicability.

Keywords: deep learnin, tiny pest detection