Detection of Densely Aggregated Wheat Rust Spores in Complex Background Microscopic Images

To solve the issue of inaccurate detection of summer spores of wheat stripe rust due to complex backgrounds, small targets, and dense adherence among individual spores, an improved model based on YOLOv7 is proposed. This work involves creating a summer spore sample dataset to establish a complex environmental data set for simulated collection. The addition of the CBAM attention mechanism helps elevate the model’s precision in detecting spores. Moreover, the ASFF adaptive feature fusion mechanism has been improved to preserve more information of small target features, thus reducing omissions. The introduction of the Wise-IoU loss function enhances the quality of detection anchor boxes and improves the model’s localization performance. Trials indicate that the revised model has achieved an mean average precision (mAP) of $98.1\%$ at the IOU threshold of 0.5. Compared to other models, both the average accuracy of detection and the speed of detection have been significantly boosted. The outcomes demonstrate that the enhanced model can accurately and precisely detect wheat stripe rust summer spores.