Research On Lightweight And Real-time Remote Sensing Object Detection Based On Attention Mechanism And Feature Fusion

With the rapid development of deep learning and remote sensing information technology,the performance of optical remote sensing image object detection has been significantly improved and has great application value in military and civilian fields.However,optical remote sensing images usual y suffered in complicated background,large object scale variations,and small objects,which makes it more difficult to detect the object in an optical remote sensing image.In addition,although the deep learning-based object detection algorithms have higher accuracy,these models are complex with large parameters and high calculation,which requires high hardware resources,making such models difficult to deploy on the platform with limited computing resources such as satellite and embedded devices.Therefore,the current optical remote sensing image object detection method still has a wide space to improve,the research on better and more efficient object detection algorithm has important research significance and application value.According to the above problems,this paper conducts research on the basis of the real-time object detection algorithm SSD to improve the performance in optical remote sensing images and lightened the feature extraction network to achieve high accuracy,lightweight and realtime object detection.The main contributions are as follows:(1)Aiming at the difficulties of small object size,complex background,and large object scale variations in optical remote sensing image,an improved method AF-SSD based on SSD is proposed.Firstly,a multi-layer feature fusion structure is designed to introduce the semantic information of the high-level features into the shalow features to enhance the semantic information of the shalow features.Next,a dual-path attention module is introduced to screen the feature information by spatial attention and channel attention to suppress the background noise and highlight the key feature,and the feature representation ability is further enhanced through a multi-scale receptive field module.Then,the loss function of SSD is improved to solve the problem of positive and negative samples imbalance.Finally,the effectiveness of the proposed method and each module are verified by a number of comparative experiments.(2)In order to solve the problems of a large number of parameters,a large amount of computation and large storage space of the traditional convolutional neural network,combined with lightweight convolutional neural network Shuffle Netv2 and SSD,a lightweight real-time remote sensing object detection SSDLite is proposed that is easy to deploy and apply.Moreover,the basic unit of Shuffle Netv2 network is optimized for the problem of limited receptive field and insufficient information representation of shallow features of Shuffle Netv2.First,the 3×3 depth convolution in the basic unit is replaced by two 1×5 and 5×1 depth convolutions to expand the receptive field of the model,and alleviate the problem of parameter increase caused by using large-scale convolutions.At the same time,in order to make the accuracy of the model not affected by the batch size during training,group normalization is used instead of batch normalization in the basic unit,so that the model can be trained in smaler batches size without affecting the accuracy of the model.Experimental results show that SSDLite can achieve real-time detection on CPU with fewer parameters and storage space,and lower computational complexity on the premise of ensuring accuracy.Finally,a comprehensive experiment is performed on two remote sensing datasets by combining the lightweight feature extraction network and AF-SSD.The results show that the proposed method has a higher accuracy,and greatly compresses the storage space and computational under the condition of less precision loss.The effectiveness and feasibility of the proposed model are verified by several comparative experiments.