Research On Light-weight Remote Sensing Image Object Detection Method Based On Meta-learning And Channel Pruning

With the rise of the era of artificial intelligence and big data,remote sensing has become ubiquitous.Optical remote sensing image object detection is a new technology emerging with the development of remote sensing.It is one of the ways to understand optical remote sensing images,and has important research significance and application value.In recent years,with the rapid development of machine learning theory and deep learning algorithms,object detection algorithms based on deep convolutional neural networks have been greatly developed and applied in many fields.However,the current object detection algorithm still has many limitations in practical applications,such as low detection accuracy for small targets,poor real-time object detection,and poor robustness in the face of complex scenes.Optical remote sensing image is precisely a small target,large scene,high-resolution data.How to quickly and accurately obtain target information from optical remote sensing images with complex background interference,arbitrary direction of target features and a large number of small targets has increasingly become one of the hot spots and difficulties in the development of object detection technology today.At the same time,with the huge performance improvement of the deep convolutional neural network,the object detection model has become deeper and wider,which greatly increases the parameters and computational complexity of the model.Redundant object detection models not only increase the hardware overhead during training,but also increase the difficulty of deploying the models on embedded devices with limited computing and storage resources.Therefore,it is of great significance to study efficient,accurate and robust optical remote sensing image object detection algorithms,and has a very broad application prospects.In this thesis,aiming at the problems that the object detection algorithm is difficult to detect on optical remote sensing images and the complexity of the object detection algorithm is high,there are two different ways to improve the detection accuracy of the object detection algorithm on the optical remote sensing image and reduce the computational complexity of the object detection algorithm.Departing from the mission,in-depth research was carried out.Specific research contents and innovations include:1.A deformable convolutional feature pyramid network is proposed.The traditional feature pyramid network structure has the problem of semantic conflict when merging feature maps of different scales,which leads to the suppression or even disappearance of the features of small objects in the feature maps after fusion,which in turn affects the detection accuracy of the object detection model.In addition,there are a large number of target objects with large differences in shape and scale and arbitrary directions in optical remote sensing images,and the number of targets is dominated by small targets,which further increases the detection difficulty of the object detection model.Aiming at the above problems,this thesis proposes an optical remote sensing image object detection method based on deformable convolutional feature pyramid network,which is used to improve the accuracy of the object detection network model of optical remote sensing images.The deformable convolution feature pyramid network mainly includes two parts: the feature enhancement module and the feature refinement module.The feature enhancement module uses deformable convolution to enhance the feature information of the target before the horizontal connection of the feature pyramid network,and the feature refinement module uses the channel attention mechanism to fuse the fusion.After the target features are refined,the information loss of the target features in the feature map fusion is reduced.At the same time,in order to further extract complex target features in optical remote sensing images,part of the twodimensional convolution in the backbone network is replaced with deformable convolution.The experimental results of this thesis on NWPU VHR-10 and DOTA v-1.0 optical remote sensing image datasets verify the effectiveness of this work.2.A lightweight optical remote sensing image object detection method based on meta-learning automated channel pruning is proposed.The object detection model based on deep learning has the disadvantages of redundant parameters and high computational complexity,and the model will consume a lot of computational and storage costs during training and inference.Traditional channel pruning methods rely on specific constraints provided by experts and scholars in the process of network pruning,which increases the labor cost of network compression and acceleration.Although the current advanced automatic channel pruning methods can automatically search for the optimal pruned network,most of these methods rely on reinforcement learning to learn the pruning strategy,which will introduce additional parameters and increase the computational complexity of the pruning method.Aiming at the above problems,this thesis proposes an automatic channel pruning method based on metalearning,which is used to efficiently search for the best pruned network structure.Based on the law of stacking block structures in most of the current redundant networks,this method builds a more streamlined network structure encoding method,trains the weight generator network to generate a pruned network for each network structure encoding vector,and uses an evolutionary algorithm to search the optimal pruned network structure.The optimal pruning network obtained from the search is transferred to the task of optical remote sensing image object detection,thereby reducing the amount of weight parameters and computational complexity of the object detection model.The effectiveness of our pruning method is verified by pruning experiments on CIFAR-10 and ImageNet datasets.Through the object detection experiments on NWPU VHR-10 and DOTA v-1.0 optical remote sensing image datasets,it is verified that our pruning method has certain transfer learning ability,which can effectively reduce the amount of model weight parameters and computational complexity while ensuring the detection accuracy of the object detection model.