In this paper, we propose an improved target detection algorithm based on YOLOX with attention mechanism for infrared aerial photography, which is used to solve the problem that small targets are easily lost in the environment with dense targets and complex background. First of all, a new channel attention module named SA-NAM is introduced. This makes it possible to utilize the contribution factor of weights to optimize attentional mechanism and SA function to enhance the feature extraction ability in network. Target distribution in real scenarios has long tail distribution, we also adopted Mosaic and Cutmix data enhancement methods and Focal loss function to category imbalance. Moreover, in order to fill the gap of infrared small target data set of aerial views in current target detection field, an infrared aerial small target data set containing human, motor vehicle and non-motor vehicle targets in various scenarios is constructed. Experiments show that the mAP of proposed algorithm on our data set has reached 63.7%, and the detection accuracy is improved by 6.2% compared with baseline YOLOX, which alleviates the misdetection of infrared small target has been ameliorated to a certain extent.
|