Aerial Object Detection Based On Deep Convolutional Neural Network

In recent years,with large-scale image datasets in the field of deep learning been disclosured on public,an increasingnumber of object detection algorithms based on deep learning have been proposed.Many experiments have been done on these datasets,where excellent performance is shown on them.However,due to the scale differences of the detected targets in the image,the situation of being occluded,and the background and light changes are easily affected in the detection process..As a result,detection results still have some shortcomings such as high recognition false alarm rate and high missed detection rate.Thus,designing a robust detection algorithm which is suitable for different application scenarios is still challenging.An algorithm which can effectively solve the problem of missed detection and false alarm in aerial images is investigated in this paper.The main works are listed as follows:(1)In view of the problem that quite a number of objects in aerial images are blocked or low in pixels,which makes it difficult to be identified effectively,an improved algorithm has been designed.The designed algorithm strikes a balance between speed and detection accuracy.In the SSD network,not all channels of the feature map contain useful detection information,so we introduce a channel attention mechanism.This mechanism can help the network to select the channels that contain useful information in the feature map,so as to enhance the discriminative ability of the feature map.Compared with the traditional attention model,detection and recognition effect of this mechanism is better.From experimental results,it can be seenthat after incorporating the channel attention mechanism into the SSD network,the performance has been significantly improved,especially for some occluded situations,this improved algorithm exhibits excellent detection capabilities.(2)In view of the large difference in the target scale in image,the anchor box is difficulty to cover targets of all sizes,a multi-branch network framework is proposed.This framework can effectively solve the problem of excessive difference in image target scale,especially in some complex scenes.Although the speed of this improvement to the framework is slightly slower than some single-branch networks,the detection accuracy is effectively improved.A large number of experiments have been done to evaluate the performance.It follows from experiment results that the multi-branch target detection algorithm proposed in this paper effectively improves the detection accuracy.(3)In the convolutional neural network,although the low-level high resolution feature map retains the spatial location information of the image,it loses the semantic information of these objects.And although the high-level feature maps in the network undergo deeper abstraction contain strong semantic information,it loses the spatial location information of the target.To overcome the above problem,a feature fusion mechanism is introduced to fuse high-level feature maps and low-level feature maps effectively.The network can retain all kinds of information in the image which is conducive to improving the detection accuracy by this mechanism.It can be seen from experiment results that the proposed algorithm has achieved great improvement in image detection accuracy.