Aiming at the problems of dense target distribution, complex background and many small targets in remote sensing image target detection, this paper improves Vovnet, adds a CoT global feature extraction module to the feature extraction backbone, which cooperates with cross-perspective feature extraction, and retains the perspective information of receptive field on multiple scales to extract the context information of targets for different scales and enhance visual representation. At the same time, a context information fusion module, namely MSSFPN, is designed based on FPN, which is built on the deep feature map. The image features are fused at the scale level to enhance the feature representation of the target. The depth hyperparametric convolution layer is introduced for prediction, and independent weights are adopted for the feature map of each channel so that the network can adapt to the image features extracted in different scales to improve detection accuracy. Compared with the original Vovnet algorithm, the mAP of the improved algorithm in the public Visdrone dataset is improved by 6.80 percentage points, which is also superior to other target detection algorithms. Experimental results further verify the accuracy and effectiveness of the improved algorithm in target detection in remote sensing images.