Local Classification Guiding Attention for Malania oleifera Recognition

Malania oleifera is a rare and endangered tree species native to natural and mixed forests, where it coexists with various other plant species. The interlaced tree crowns present significant challenges for conventional detection methods. This study proposes a novel approach to detect Malania oleifera tree crowns in complex environments. The method utilizes a Local Classification Guiding (LCG) attention mechanism and a pixelbased intersection over union (PIoU) loss function to improve the detection performance. The LCG module segments the feature map into small patches to enhance the contribution of pixels within the local region, and employs a classification guiding attention module to establish relationships between the pixels and classes. The PIoU loss function calculates the loss at the pixellevel instead of the traditional box-level loss function to improve the detection precision of crown shape. The experimental results show that the proposed method achieves superior performance compared to traditional mask detection methods such as Mask RCNN, CondInst and YOLOv8. It detects more targets and delineates tree crown masks more accurately. The LCG module increases precision by 0.2, recall by 0.05 and F1-score by 0.12. These results indicate that the LCG effectively improves the tree crown detection. The PIoU loss function significantly improves detection accuracy, with the pixel level IoU improved from 0.67 to 0.83. By integrating the LCG module and the PIoU loss function, the average precision reached 0.979, average recall reached 0.836, and average F1-score reached 0.872. This method provides a new attention mechanism for object detection and a pixel level loss function for detecting irregularly shaped objects