Self-Supervised Incremental Learning for Insect Segmentation and Counting on Resource-Constrained Devices

Lack of proper real-world data is a fundamental issue of deployed edge AI systems, which can lead to inaccurate AI model performance and, eventually, to system failures. Incremental learning (IL), in which the model is retrained on newly incoming data, can be used to overcome this issue. This study proposes a novel, comprehensive self-supervised IL framework for resource-constrained image analysis applications. It focuses on insect segmentation and counting, fundamental operations for edge Internet of Things (IoT) devices deployed in pest control applications. The framework distributes actions between edge nodes, relying on low-power microcontrollers (MCU), and an edge server, implemented on a more powerful processor device, such as a Raspberry Pi (RPi). Segmentation and counting tasks are performed on edge nodes using a lightweight deep-learning (DL) model, while the edge server autonomously handles image reconstruction, mask generation, and pseudolabeling to incrementally retrain the AI model. This architectural concept eliminates the need for continuous manual labeling and enables continuous retraining on the edge without transferring large amounts of data to the cloud. Through a series of experiments, we demonstrate accuracy comparable to the manual annotation of images by humans, along with approximately a 10% improvement in segmentation and a threefold reduction in counting error over the base model. In addition, the results reveal low retraining time (363.3 s) and power consumption (1.2 A at 5.1 V) on an embedded edge server, demonstrating that the framework is extremely suitable for agricultural setups lacking a fixed power supply or network connection, especially given that retraining is only needed occasionally.