Research On UAV Image Small Object Detection Algorithm Based On Feature Enhancement

In recent years,with the development of convolutional neural networks,the rapid development in the field of object detection has been promoted.However,the detection of small objects is still one of the difficult problems in the field of object detection.UAVs are widely used in disaster relief,traffic surveillance,video shooting,disaster surveillance,and industrial object detection.The images captured by UAVs contain a large number of small objects,and the small size of target objects,less feature information,and dense object detection in such images lead to the unsatisfactory detection effect of mainstream object detection algorithms for small objects in UAV images.Therefore,the problem of small object detection in UAV images needs to be further studied.This chapter studies the improvement based on YOLOv5 and YOLOX object detection models,and uses feature enhancement technology to improve the detection effect of small objects,the specific research is as follows:(1)The diminutive size of the object detection in UAV image and the insufficient amount of feature information it contains pose a formidable challenge to extant detection algorithms,thereby resulting in a suboptimal detection outcome.To surmount this predicament,A UAV image small object detection algorithm based on feature fusion and attention mechanism is designed.a multi-head attention mechanism is incorporated into the YOLOv5 backbone network in order to seamlessly integrate global feature information.As the network depth increases,the model tends to accentuate high-level semantic information at the expense of underlying detailed texture features that are vital for the detection of small objects.To counteract this,a shallow feature enhancement module is devised to acquire underlying feature information and augment small objects feature information.Furthermore,a multi-level feature fusion module is crafted to amalgamate feature information from different layers,thus enabling the network to dynamically adjust the weights of each output detection layer.An analysis of the experimental outcomes demonstrates that the mean average precision of the proposed algorithm,as applied to the publicly available Vis Drone2021 dataset,attains a level of 45.7%,representing a 3.1%enhancement over the baseline YOLOv5 algorithm.Furthermore,the detection speed for high-resolution images,at 41 frames per second,satisfies the requirement for real-time performance and exhibits a noteworthy improvement in detection accuracy when compared to other prevalent methods.(2)Aiming at the problems of difficult detection of dense objects and loss of small objects feature information in UAV images,a small object detection algorithm based on feature enhancement of YOLOX UAV images is designed.Firstly,in order to enhance the learning of the underlying feature information,attention mechanism is combined with residual module in the YOLOX backbone network,the feature learning is performed by the residual module,and the attention mechanism is used to enhance the image feature information.This module enables the model to effectively learn image feature information.Secondly,due to the difficulty of detecting dense objects in UAV images,the varifocal loss function is introduced to realize IOU-aware classification score regression,so as to improve the detection effect of dense objects.Finally,the weight smoothing of the model is optimized,and the improved exponential moving average algorithm can make the model converge quickly in the early stage of training,and the convergence of the model in the later stage of training will be more stable.The experimental results show that the proposed algorithm has a good effect on the average mean accuracy of the baseline algorithm on the public dataset Vis Drone2021,and has better performance than other mainstream algorithms.