|
Based on the ability of YOLOv5 algorithm for maize weed image recognition, an improved model is proposed to promote the precision of the model for weed identification by adding the attention mechanism ACmix. The backbone network module has been enhanced by incorporating the ACmix module to substitute the 3×3 convolution kernel, while the detection head part has been upgraded through replacement of the 3×3 convolution with the CBAM attention mechanism. The detection accuracy of the model is improved by learning the feature map information at multiple target scales simultaneously. To explore the performance of the model, we collected large number of corn and weed images, built our own dataset and labeled it using the Labelimg tool. Regarding the preprocessing of data, techniques including data augmentation were utilized to improve the model's ability to generalize. We also performed experimental validation using the dataset, and the results show that our improved model achieves significant improvements in both detection and classification tasks of corn weed images compared to the original YOLOv5 algorithm. Finally, we conducted comparative experiments on the ACmix attention mechanism and compared it with other commonly used attention mechanisms. Through training experiments on the established dataset, it is found that the improved model achieves 94.3%precision and 82.5 %recall. Compared with the original YOLOv5 model, the precision improved by 3.0% and recall improved by 1.5%; compared with YOLOv4, the precision improved by 4.6% and recall improved by 1.8%. The results show that the ACmix attention mechanism has advantages in improving model performance. In summary, this paper proposes an improved method based on YOLOv5 deep learning neural network to improve the accuracy of maize weed image recognition by adding the attention mechanism ACmix, which has practical application value. A corn seedling weed dataset was established to provide a data base for subsequent research on corn weeds.
|
