In recent years, target detection, as an important branch of computer vision technology, has been widely applied in fields such as medicine, military affairs, and urban rail transit. As satellite and remote sensing technologies advance, images obtained using these technologies contain abundant information. This makes it crucial to conduct automatic target detection and understanding of these images. However, due to the random directions and dense distribution of targets in remote sensing images, conventional methods are prone to lead to missing or incorrect detection. In response, this study proposes a multi-convolution kernel feature combination-based adaptive region proposal network (MFCARPN) algorithm for multi-directional detection. This algorithm introduces multiple convolution kernel features for target extraction. The weight parameters of these convolution kernel features can be determined through adaptive learning according to the differences between the targets, yielding the characteristic patterns that match better with targets. Meanwhile, in combination with the original features of the targets, the parameters of the classification and regression model vary dynamically according to the difference between targets. Thus, the RPN’s adaptive ability can be improved. The experimental results indicate that the mAP of the standard dataset DOTA reached up to 75.52%, which is 0.5 percentages higher than that of the baseline algorithm GV. Therefore, the MFCARPN algorithm proposed in this study proves effective.