Recently, many approaches based on deep convolutional neural networks (CNNs) for object detection have showed better performance compared with traditional approaches. Since objects in high-resolution aerial images are usually very small with complex backgrounds, the performance of deep CNN-based approaches on object detection in high-resolution aerial images is still limited. In addition, with the constraint of the amount of memory on graphics processing units (GPUs), current state-of-the-art CNN architectures cannot directly process high-resolution aerial images. This paper proposes an improved deep CNN-based framework for object detection in high-resolution aerial images. To detect objects at different scales, image pyramid with different sizes is generated from single input image by down-sampling the original high-resolution input image. Each image level in the image pyramid can be used to detect objects at a different scale. For each image in the image pyramid, small patches with small fixed-size are generated. Each patch is then fed to the detection network, which is based on SSD framework with VGG-16 architecture as the base network, to generate patch detection results. Patch detection results are then projected to the image pyramid at the original scale to form image detection results. Finally, Non-Maximum Suppression (NMS) algorithm is adopted after image detection results to create final detection results. Experimental results on public datasets with highresolution aerial images show that the proposed approach is very simple and efficient while achieving nearly as detection accuracy as recent state-of-the-art methods.