Lightweight Apple disease detection algorithm based on improved YOLOv11s

This study addresses key technical challenges in detecting small-target apple diseases such as anthracnose and Marsson’s spots within complex orchard environments, including high background complexity, feature blurring caused by uneven illumination, and loss of detailed features during conventional downsampling. We propose a lightweight and efficient detection model, AS-YOLOv11s, which first optimizes the Backbone network by introducing an adaptive downsampling module (ADown) to replace traditional convolution. This modification enhances the model’s sensitivity to small-target diseases by preserving critical contour and texture features while eliminating redundant information. Additionally, a spatial-channel synergistic attention mechanism is integrated into the key layers of the Neck network, enabling multiscale lesion feature extraction through coordinated attention across spatial and channel dimensions. Experiments were conducted on a self-constructed apple disease dataset, partitioned into training, validation, and test sets at an 8:1:1 ratio. Compared to the original YOLOv11s model, AS-YOLOv11s demonstrated notable improvements: precision for anthracnose detection increased by 3.4%, recall by 0.8%; recall for Marsson’s spots detection improved by 10.0%; and the mean average precision at 0.5 IoU ([email protected]) increased substantially by 17.0%. The [email protected] for healthy apple detection also improved by 0.6%. Concurrently, the model’s parameter count and computational complexity were reduced to 79.5% and 77.4% of the original model, respectively, with a final model size of only 14.6 MB, demonstrating advantages in both detection accuracy and model compactness. These results validate the effectiveness of the proposed algorithm for efficient apple disease detection in complex orchard scenarios, providing robust theoretical and technical support for the practical deployment of intelligent orchard inspection systems. 1. Replacing traditional convolutional downsampling with a lightweight ADown module, which effectively preserves subtle lesion textures through parallel pooling path fusion while mitigating feature degradation in deep networks. 2. Designing a Spatial-Channel Synergistic Attention (SCSA) mechanism that collaboratively optimizes spatial and channel dimensions to autonomously capture multi-scale lesion features and allocate weights. 3. The AS-YOLOv11s model reduces parameters, computational load, and memory footprint to 79.5%, 77.4%, and 14.6 MB respectively compared to the original model, delivering both high accuracy and lightweight advantages.