A Novel AI-Driven Approach for Synthetic Data Generation in Camouflaged Insect and Pest Detection

Data scarcity of agricultural computer vision data in the area of Camouflaged insect and pest detection remains a challenging problem, despite recent advancements in Camouflaged Object Detection (COD) and transformer based architectures. To address this problem, we introduced a novel AI-driven synthetic image generation methodology that leverages computer vision models, Large Language Models(LLMs) and Generative AI to create challenging synthetic camouflaged images of insects and pests. This synthetic data helps enhance the generalisation capabilities of COD models.

The proposed methodology includes an iterative process involving multiple AI agents performing (1)extraction of pest-crop knowledge using LLMs, (2) collecting pest and plant images in an automated way, (3) separating out the foreground pest from background, (4) generating synthetic images with varied camouflage parameters, (5) evaluating using existing detection models. (6) incorporating images that are not detected into the training dataset, enabling iterative dataset enhancement (7) retraining the model. This approach accelerates the development of intelligent, sustainable and scalable agri-tech solutions by providing critical training data that is inexhaustible and relevant.

We have experimentally validated the proposed iterative methodology on the detection accuracy for camouflaged hornworm pests on tomato plants. In our ablation studies, we have seen that the baseline mAP50 for models trained on web-scraped images was 0.170 when tested with camouflaged images, whereas the models trained with images generated using our methodology have shown a mAP50 value of 0.931 when tested with the same set of camouflaged images. Using the Gradient-weighted Class Activation Mapping (Grad-CAM), we provided evidence on how the iterative blending process in our methodology is able to regularise the model and focus on the biological anatomy and texture of the pests instead of just the edges and boundaries, proving that the methodology not only solves the data scarcity of camouflaged pest images but also improves the robustness of models in complex environments.