Small Objects Detection in Satellite Images Using Deep Learning

Using the deep Convolution Neural Networks (CNNs) for Object detection in satellite images accomplish promising results, especially for large objects. While Small objects detection in the same spatial resolution images does not attain the same results. For instance, vehicle detection in high-resolution satellite images, the targeted object maybe existed in an area that does not exceed 15 square pixels, which will not make a sufficient effect in the deeper layers. In addition; the interfering with the surrounding background, noise effect, the neighboring object's shadows, and various vehicle colors. In the proposed paper, an analysis study is performed to evaluate the effect of changing the object size on the detection results. A separate resampling algorithm is applied to the input test images to change its size - bear in mind the built-in detection model resampling layer-, which results in changing the object size, and accordingly extends the object impact in deep layers. Through Transfer Learning, the Faster R-CNN pre-trained object detection model with Inception-V2is applied to submeter satellite images and passenger vehicles as the target objects. The Experimental results show the change in detection accuracy with the change of the object size.