Research On Aerial Image Target Detection Technology Based On Convolutional Neural Network

With the rise of deep learning,computer vision has been greatly developed,and object detection as an important research field in computer vision has attracted many scientific researchers to join it.Target detection is widely used,and its main task is to identify the target on the input image and give the corresponding position information.Nowadays,data presents a blowout development,and deep learning requires a large amount of data for training and learning.Based on this large environment,superior performance target detection algorithms are constantly being proposed by researchers.Target detection can be used in many scenarios,and its basic requirement is to ensure the accuracy of detecting targets.However,due to sampling equipment,sampling environment,target scale and other reasons,its detection accuracy is often not guaranteed due to these external factors.Moreover,in some scene tasks,such as drone detection,it is necessary to ensure the real-time detection.Some algorithms with superior detection performance require a large amount of calculation.While the detection accuracy is guaranteed,the real-time performance cannot be guaranteed and cannot be applied to some In specific scenarios.Therefore,in the field of target detection,there are still many problems waiting to be solved.This article mainly studies the target detection of aerial images.In aerial images,small targets are densely distributed and have directions,and their size is too small.With the deepening of the convolutional neural network,small targets are expressed in deep features.The semantic information of is basically ignored,while the shallow features can better express the semantic information of small targets.In response to this problem,inspired by the idea of fusion feature layer,the basic network VGG16 is improved,and its deep features are merged with shallow features by upsampling.The obtained fusion features can effectively express the small target in the final feature map.In order to solve the problem that the detection result does not match the real target,this paper adds a fully connected layer to the proposed network,and through learning the relative offset between the real target,and finally output the rotated Ro I,perform feature extraction on the rotated Ro I,pool Classification and regression.This design effectively solves the problem of mismatch of horizontal bounding boxes in aerial image detection.As the neural network deepens,the network model is prone to overfitting.In order to prevent this problem,this article adopts some optimization strategies,such as expanding the data set through rotation,translation,and clipping,batch standardization,and L2 regularization operation Wait.The model proposed in this paper effectively solves the problem of mismatch in aerial image detection without adding too many computing resources,improves the detection accuracy,and prevents the model from overfitting.It is obtained on the public data set DOTA The experimental results verify the effectiveness of the model.