Aircraft Detection Based on Saliency Map and Convolution Neural Network
Yuwei Sun, Nagul Cooharojananone, Hideya Ochiai · 2019
Detection of the aircraft from the remote sensing images attracts many attentions. Not only the aircraft taking off and landing but also the aircraft landing in the airport need to be monitored. This paper proposes a saliency-based CNN (convolutional neural network) method for aircraft detection in the remote sensing images. Many researches of aircraft detection access to a method of R-CNN (region convolutional neural network), which generates two thousand or more of proposal regions as candidates of the aircraft. The researches of using a saliency map to extract the objects from the images are still not so many. We adapt a series of preprocessing as transforming to a gray scale image, deleting the noise, binarizing by threshold, the closing operation, and floodfill to generate a saliency map where the contours of all objects are reinforced. Compared to R-CNN, the using of a saliency map greatly reduces the number of proposal regions, thus improving the efficiency. After this, we find the contours of objects and the minimum rectangles enclosing objects to extract proposed regions. Then, we prepare a negative dataset consisting of different types of backgrounds including the land, grass land and concrete. And for a positive dataset, we prepare images of aircraft in different angles and different sizes. To add more variation data into the dataset, we also use methods of augmentation like elastic distortion and perspective transforms. Then, we train a CNN model using the prepared dataset, which can tell whether a proposed region contains aircraft, after training. At last, we use a method named recall to evaluate the performance of the scheme, attaining an average recall of 0.7217, which shows the rate of aircraft successfully detected in all aircraft.