Abstract: Due to the large difference in the scale of the object, the arbitrary direction and the dense distribution of the object in the remote sensing image, the existing detection methods rarely pay direct attention to the dense edge information and the object cannot obtain a suitable receptive field, so it is difficult to have good detection results in remote sensing detection. In order to solve the above problems, this paper proposed a multi-scale feature enhancement network based on large kernel convolution and dense object refinement（LKCSFP-NET） for remote sensing image detection. Firstly, the network based on SKNET added a cavity convolution to form a large kernel convolution block（LKB） to obtain the best sensitivity field for small targets and improve the adaptability and accuracy of the network to multiple scales. Secondly, on the basis of FPN, the centralized spatial feature pyramid CSFP module was added to solve the problem of low detection efficiency of remote sensing images due to dense object distribution and complex detection background by combining global semantic information with local semantic information. The experimental results show that on the DOTA and HRSC2016 public datasets, the average detection accuracy of the proposed algorithm on the two datasets is 74. 90% and 96. 60%, respectively, which is 1. 36and 0. 63 percentage points higher than that of the baseline network, which is better than most existing models. The proposed LKCSFP-NET has stable performance in the two public datasets, and has good detection results for small objects and densely arranged objects, which is higher than the detection accuracy of most existing models, and can be well applied to the detection of remote sensing objects.