作者:\n\t戴天虹，刘超\n

Authors:\n\tDAI Tianhong,LIU Chao\n
摘要:\n\t针对红外相机图像中雪豹存在难检测、难辨认等问题，提出了一个基于域迁移和新型注意力机制的EfficientDet雪豹检测算法。该算法首先采用图像增强来扩充训练样本提高模型的鲁棒性，并添加非雪豹图像优化数据集结构；其次，使用生成式对抗网络对夜间红外灰度图像进行域迁移，融合图像迁移前后的预测结果解决夜间红外图像目标识别能力弱的问题；最后通过在主干特征提取网络中加入注意力机制细化特征图来提高RGB和红外图像中雪豹的视觉显著性，并且改进特征融合网络结构整合更多有效信息。分析对比实验和消融实验的结果表明，该方法相比Faster-RCNN、YOLOv3和SSD的检测速度和精度更好，改进后的算法平均精确度为97.4%且检测速度可达19张/s,该检测模型更适合定位识别雪豹RGB与红外图像。\n

Abstract:\n\tIn view of the difficulty of snow leopard detection and recognition in infrared camera images, a snow leopard detection algorithm is proposed based on EfficientDet, which combines domain migration and new attention mechanism.Firstly, the algorithm adopts image enhancement to expand the training sample and adds non-snow leopard images to optimize the dataset structure to improve the robustness of the model.Secondly, the Generative Adversarial Networks is used to transfer the domain of night infrared gray image, and the image prediction results before and after migration are fused to solve the problem of weak target recognition ability in infrared image.Finally, the attention mechanism is added to the backbone feature extraction network to refine the feature map to improve the visual significance of snow leopard in RGB and infrared images, and the improved feature fusion network structure to integrate more effective information.The results of the contrast experiment and the ablation experiment show that the proposed method has better detection speed and accuracy than Faster-RCNN，YOLOv3 and SSD. The improved algorithm has an average accuracy of 97.4% and detection speed of 19 images/s.This detection model is more suitable for locating and recognizing snow leopard RGB and infrared images.\n


PDF全文下载地址:

可免费Download/下载PDF全文