Abstract:

A lightweight detection algorithm (AGLU-YOLO) is proposed to solve the problems of insufficient accuracy, small lesions and complex orchard background in citrus leaf disease detection. This method fuses AdditiveBlock and Convolutional Gated Linear Unit (CGLU) in the C3k2 module of the backbone network to form the C3k2_AdditiveBlock_CGLU module. The former enhances long-range dependence and global context representation by additive modeling, and the latter realizes conditional gating by depthwise separable 3 × 3 convolution and point convolution to suppress false activation caused by complex texture and enhance small-scale lesion response. At the same time, the AFCA attention mechanism is added in the feature fusion stage to improve cross-layer semantic interaction and multi-scale robustness. Secondly, in order to meet the needs of edge deployment, the LAMP hierarchical importance pruning algorithm is used to jointly compress the channel / level, and a lightweight fine-tuning is performed to restore the accuracy; then the model is exported to ONNX and operator fusion and low-precision inference optimization are implemented through TensorRT to achieve real-time detection with low latency and high throughput. Through experimental verification on the self-made dataset, AGLU-YOLO improves the Precision, Recall and mAP @0.5 indexes by 3.1, 4.4 and 5.1 percentage points, respectively, compared with the baseline YOLOv11n, and takes into account the volume and recognition speed, showing stronger robustness and multi-scale lesion adaptability, which can effectively meet the application requirements of rapid and accurate identification of citrus leaf disease.

Key words: citrus leaf disease detection, small target detection, YOLOv11 algorithm, deep learning, CGLU