| Object detection technology is one of the basic tasks of computer vision.Object detection technology for remote sensing images is a specific field of object detection technology.In recent years,with the emergence of a number of remote sensing image datasets and the extensive use of depth learning in computer vision,the remote sensing image object detection technology has developed rapidly,and its main steps include feature extraction,feature fusion and detection.Although many scholars have done a lot of research on these three steps,there are still some problems to be solved urgently in the current remote sensing image object detection technology,such as the poor accuracy of the feature extraction network or the large computational complexity,the loss of a lot of information in the feature fusion process,and the poor adaptability of the detection module to the remote sensing object.This paper has carried out systematic research on the above problems,and the main work is as follows:1.Aiming at the problem that the feature extraction module is difficult to achieve the desired detection effect with low computational complexity,a remote sensing image object detection algorithm based on ConvNeXt network is implemented.First of all,the algorithm uses ConvNeXt network to extract the feature information in remote sensing images.Compared with the current feature extraction network based on Transformer,ConvNeXt network maintains a higher detection accuracy with a lower computational complexity.Secondly,in the feature fusion module,a lightweight attention mechanism is introduced to further filter the extracted features,which can assist the ConvNeXt network to achieve better detection accuracy without increasing the complexity of the algorithm.The experimental results show that the detection accuracy of this algorithm on DIOR,NWPU VHR-10,and RSOD datasets has been improved by 4.9%,1.7%,and 2.5% compared to the baseline,respectively.2.Aiming at the problem that object feature information is lost in the process of feature fusion,a feature enhanced object detection algorithm for remote sensing images is proposed.First of all,the algorithm introduces a residual feature enhancement strategy in the feature fusion process.This strategy processes the features output at the highest level of the feature extraction network and then performs the dimensionality reduction operation of the feature fusion module,which reduces the feature information loss rate in the feature fusion process.Second,to encourage the network to pay closer attention to object features,an improved channel attention mechanism is added to the feature fusion module.At the same time,the proposed attention mechanism is compared with other attention mechanisms to prove the effectiveness of the proposed attention mechanism.The experimental findings demonstrate that the algorithm’s detection accuracy on the DOTA-v1.0 dataset and the HRSC2016 datasets is 1.43% and 0.28% higher than the baseline,respectively.And on DIOR,NWPU VHR-10 and RSOD datasets,it has been improved by 1.1%,1.2%,and 1.2%,respectively.3.Aiming at the problem that the detection module is difficult to adapt to objects with significantly different proportions and arbitrary rotation angles in remote sensing images,which affects the detection accuracy,an object detection algorithm based on multidimensional attention for remote sensing images is proposed.First,an attention module is introduced into the detection module,which is cascaded by multi-dimensional attention mechanism.The attention module alleviates the problem of increasing the detection difficulty due to the large shape difference between the objects in the remote sensing image and the large change of rotation angle and position.Secondly,the pooling operation in the channel attention mechanism is replaced by discrete cosine transform and its activation function is modified to form an improved frequency channel attention mechanism.This mechanism can strengthen the object features output by the feature extraction module,and the auxiliary detection module can output more accurate detection results.Compared to the baseline,the detection accuracy of this algorithm has been improved by 1.46%,0.4%,and 1.8% on DIOR,NWPU VHR-10,and RSOD datasets,respectively. |
