基于改进Shuffle-RetinaNet的红外车辆检测算法

Infrared Vehicle Detection Algorithm Based on Improved Shuffle-RetinaNet

针对当前红外场景下多尺度车辆检测精度欠佳且算法模型复杂度高的问题,提出了基于Shuffle-RetinaNet的红外车辆检测算法。该算法以RetinaNet网络为基础,并选用ShuffleNetV2作为特征提取网络。提出双分支注意力模块,通过双分支结构和自适应融合方法增强网络对红外图像中目标关键特征的提取能力;优化特征融合网络,集成双向交叉尺度连接和快速归一化融合,增强目标多尺度特征的表达能力;设置校准因子增强分类和回归之间的任务交互,提高目标分类和定位的准确性。该算法在自建红外车辆数据集上的检测精度达到92.9%,参数量为11.74×10^(6),浮点计算量为24.35×10^(9),同时在公开红外数据集FLIR ADAS上也展现出较好的检测性能。实验结果表明:该算法具有较高的检测精度,且模型复杂度低,在红外车辆检测领域具有较高的应用价值。

In view of the low detection accuracy and high complexity of current multi-scale vehicle detection algorithms in infrared scenes,an infrared vehicle detection algorithm based on Shuffle-RetinaNet is proposed.On the basis of RetinaNet,the algorithm uses ShuffleNetV2 as the feature extraction network.A dual-branch attention module channel attention module is proposed,which adopts the dual-branch structure and adaptive fusion and enhances the ability to extract the key features of the target in infrared images.To optimize the feature fusion,the algorithm integrates cross-scale connection and fast normalized fusion in some feature layers to enhance the multi-scale feature expression.The calibration factor is set to enhance the task interaction of classification and regression,and the accuracy of target classification and locating is increased.A series of experiments are conducted on a self-built infrared vehicle dataset to verify the effectiveness of the proposed algorithm.The detection accuracy of this algorithm for the self-built vehicle dataset is 92.9%,the number of parameters is 11.74×10^(6),and the number of floating-point operations is 24.35×10^(9).The algorithm exhibits better detection performance on the public dataset FLIR ADAS.Experimental results indicate that the algorithm has advantages in detection accuracy and model complexity,giving it good application value in multi-scale vehicle detection tasks in infrared scenes.