Research On Small Object Detection Algorithm In Aerial Images Based On Deep Learning

Object detection is a technology for locating and classifying objects in images.With the rapid development of deep learning and the improvement of computing power,the object detection model based on deep learning can meet the needs of real-time detection.The size of the small objects is minor,and the effective information provided to the network is less,which makes the detection performance of the existing object detection model for small objects poor.This paper analyzes the difficulties of small object detection in aerial images,and effectively improves the detection performance of the model for small objects without affecting the detection speed.This paper mainly focuses on the following aspects:(1)The model in this paper is trained using the small object VEDAI(Vehicle detection in aerial Imagery)dataset.Aiming at the problems of small objects size and insufficient effective information for model learning,this paper performs data augmentation and clustering algorithm on small objects.Data augmentation technology refers to performing geometric transformation,color dithering and other operations on the object before training,and at the same time randomly extracting two images for random superposition to generate new samples.The clustering technology uses the K-means algorithm to cluster the small objects of the dataset,and redesigns the initial size of the anchor box.According to the above operations,the robustness of the model to small objects is better.(2)Current detectors typically exploit feature pyramid networks in multi-scale feature fusion way to improve small object detection performance.However,the salient features of large objects preventing the detection of small objects in the shallow layers,lateral connection reduce semantic information in the deep layers.Aiming at the above problems,an improved feature pyramid networks was proposed.In the shallow layers,the model is designed to erase the large objects and focus on the small object.In deep layers,the Residual Feature Augmentation model is used to reduce the information caused by the direct reduction of 2048 channels of feature maps to 256 channels.The improved feature pyramid networks effectively improves the detection accuracy of small objects.(3)Object detection based on deep learning is an efficient method.However,when the neural network extracts the features in the image,it will generate a large amount of features,and the model does not know which features are useful.In response to this problem,this paper uses the CBAM(Convolutional Block Attention Module)attention mechanism to improve the model to improve the expression ability of small object features.Experimental data show that using the CBAM attention mechanism behind the backbone network can help the model learn the areas that need to focus on small objects,and effectively improve the model’s detection accuracy for small objects.This paper conducts experiments on the VEDAI dataset,and builds a basic model for small object detection without affecting the detection speed,and the detection accuracy of small object reaches 74.0%.The improved feature pyramid and attention mechanism effectively reduce the problem of missed detection and false detection of small objects in aerial images,and the detection accuracy of the improved model reaches 77.6%.