基于多尺度特征对齐的小样本目标检测方法OA

To address the issue that the scarcity of real-world data in few-shot object detection leads to inadequate model generalization ability,we propose a few-shot object detection method based on multi-scale feature alignment.Firstly,the LaMa algorithm is employed for simulated data augmentation to generate high-fidelity simulated data and expand the training samples.Secondly,we design a salient feature extraction module that hierarchically embeds the coordinate attention mechanism to enhance feature responses in target regions and suppress complex background noise.Finally,we construct a multi-scale feature alignment module that performs hierarchical alignment processing on low-level texture,mid-level structure,and high-level semantic features.By designing a multi-scale feature alignment loss function,we enforce consistency constraints on cross-domain feature distributions.The experimental results show that compared with the benchmark algorithm,the accuracy of the proposed method has increased by 20 percentage points,the recall rate has increased by 2.8 per-centage points,and mAP@0.5 has increased by13.5 percentage points.Compared with YOLOv8,YOLOv11,SSD and Faster R-CNN object detection algorithms,the mAP@0.5 of the proposed method has increased by12.1 percentage points,8.1 percentage points,35 percentage points and30.6 percentage points,respectively.These results fully validate the effectiveness and robustness of the proposed method,providing an effective solution for the task of few-shot object detection.