Incremental YOLOv5 for Federated Learning in Cotton Pest and Disease Detection with Blockchain Sharding

Abstract In this paper, an incremental YOLOv5 model based on blockchain partitioning technology is proposed, aiming at solving the problem of spatio-temporal heterogeneity in cotton pest and disease identification in Xinjiang, as well as improving the automation, accuracy and efficiency of detection. Through the lightweight improvement and the introduction of attention mechanism with deep separable convolution, the model's inference speed and accuracy are enhanced under different computing environments. Combining federated learning and knowledge distillation techniques, the proposed IFOD framework effectively mitigates the catastrophic forgetting problem in incremental learning, reducing the amount of model parameters by 69.95% and the training time by about 60%, despite a 5.7% decrease in accuracy compared to the original model. The designed reputation evaluation and reward distribution mechanism, based on blockchain slicing, ensures high-quality contribution of data and system security. Experimental results show that the IFOD-shard framework excels in reducing the amount of model parameters and computation, increasing the detection speed, while maintaining the memory of the old target while incrementally learning the new target, and significantly reducing the training and communication costs. The reputation evaluation mechanism has excellent ability to recognize malicious nodes and ensures the fairness of reward distribution. This framework not only improves the level of intelligent identification of cotton pests and diseases, but also provides an effective solution to solve the problems of data privacy and computational resource limitations in other fields.