| With the progress of artificial intelligence,target detection technology based on deep learning is widely used in fields such as UAV inspection,autonomous driving and security.Visible small target detection in aerial photography scenes has achieved remarkable results,but detection of infrared small targets has rarely been studied due to small data sets and learning difficulties,and target detection algorithms based on visible images are difficult to fully adapt to the problem of small target detection in infrared scenes.However,infrared imaging technology has stronger penetration and anti-interference properties,which can meet the needs of specific application scenarios.Therefore,this article conducts an in-depth study on infrared small target detection in UAV aerial photography scenes and proposes an improved YOLOX algorithm,which better solves the problem of missing and false detection of small targets in complex and dense scenes of aerial photography.The main work of this article is as follows:(1)Based on the analysis of the superiority of infrared detection technology for all-weather surveillance and penetration capability,an open source aerial small target dataset containing11245 infrared images was created.The dataset images were collected by UAVs and autonomously annotated using Labelimg software for detection needs,covering a variety of attributes of target location,size and shape,providing more comprehensive and high-quality information for the training and evaluation of target detection algorithms.(2)To address the problems of small dataset size and long-tailed data distribution,two data enhancement methods,Mosaic and Cut Mix,are introduced for improvement,increasing the dataset size,balancing the number of samples,and improving the accuracy and generalization ability of the model.The value of the number of data enhancement rounds epoch is optimised,and the strategy of cancelling data enhancement for small targets below a specific pixel threshold is introduced,and the optimal solutions for epoch and small target pixel threshold are obtained through extensive experiments.(3)An improved SA-NAM(Smooth Activation-Normalization-based Attention Module)channel attention module design scheme is proposed to enhance the feature extraction capability of neural networks for weak infrared targets.Batch Normalization(BN)is introduced in this module to normalize the intermediate output values of the neural network,and a method is designed to dynamically represent the attention weights of each channel using the scaling factor of BN.The sparse weight penalty makes the computation more efficient,and the dynamic weighting allows the network to focus more on small-scale targets that are difficult to learn.In addition,the Sigmoid activation function is improved by a smoothly derivable SA(Smooth Activation)function at the origin,further improving the effectiveness and real-time performance of the network training.(4)To address the problems of positive and negative sample imbalance and insufficient ability of the Binary Cross Entropy Loss(BCE)loss function in the YOLOX algorithm to handle hard-to-classify samples,the optimal estimation method based on Focal loss is investigated,and the weight parameter and adjustment factor are introduced to adjust the positive and negative samples and the hard and simple samples respectively,so that the model can perform better in fine-grained classification.(5)Comparative experiments based on YOLOv3,YOLOv5 and YOLOX target detection algorithms are conducted to demonstrate the superiority of the improved YOLOX algorithm for infrared aerial photography small target detection,and the effectiveness of each improved strategy when applied to the model alone is verified through ablation experiments.The experiments show that the improved algorithm improves the accuracy metric mAP0.5-0.95 on the test set by 4.6 percentage points relative to the baseline model,effectively reducing the model’s miss and false detection rates. |
PDF Full Text Request