A study on an efficient citrus Huanglong disease detection algorithm based on three-channel aggregated attention

Aiming at the problems of complex and diverse field symptoms of citrus Huanglong disease (HLB), low efficiency and insufficient recognition accuracy of traditional detection methods, this study proposes an efficient detection algorithm based on improved You Only Look Once (YOLO)v8.

Firstly, a new character to float (C2f) Attention inverse residual moving block (IRMB) module is designed, which significantly enhances the model's sensitivity to tiny disease features while reducing the number of parameters by fusing the lightweight IRMB with the adaptive attention gating mechanism, and solves the problem of losing key texture information due to downsampling in the traditional C2f module. Secondly, the three-channel aggregated attention module Powerneck is proposed in the Neck section, which realizes efficient cross-scale feature interactions, effectively suppresses background noise interference, and improves robustness in complex field scenes through SimFusion_4in feature alignment, information fusion module (IFM) global context fusion, and Power channel dynamic weighting strategy. In addition, the detection head design is optimized by structural reparameterization technique to further accelerate the inference process.

The experimental results show that on the citrus dataset containing 12 diseases and two health states, the mAP50 of this model reaches 97% and the accuracy is 91.5%, which is 1.1% and 1.2% higher than that of the original YOLOv8, respectively, and the inference speed is improved by 14.6% to 370 frames per second (FPS). Comparison of the different models shows that the C2f Attention IRMB, through the mechanism of dual attention The comparison of different models shows that C2f Attention IRMB strengthens the feature expression ability through the dual-attention mechanism, and the Powerneck module reduces redundant computation through dynamic channel pruning, and the two synergistically optimize the model performance significantly. Compared with mainstream models such as YOLOv5m and YOLOv7x, this method is more advantageous in the balance of accuracy and speed, and can meet the demand of real-time detection in the field.

The algorithm provides an efficient tool for early and accurate identification of citrus Huanglong disease, which is of great practical significance for reducing pesticide misuse and improving the efficiency of orchard management, and also provides new ideas for the design of lightweight target detection models in agricultural scenarios.