Aircraft Type Recognition Based on Segmentation With Deep Convolutional Neural Networks

Aircraft Type Recognition in Remote Sensing Images Based on Feature Learning with Conditional Generative Adversarial Networks

“…Aircraft keypoint detection methods are proposed in Reference [15,16]. Specifically, they first resize images to 40 × 40, and then regress the keypoint coordinates directly using CNN models.…”

“…Then, the keypoints are transformed back to the original images and used to recognize aircraft types. However, the methods in References [15,16] use pooling layers to process the whole feature maps from the previous layers. This operation ignores the local features learned by low layers, which are important to keypoint detection tasks.…”

“…This operation ignores the local features learned by low layers, which are important to keypoint detection tasks. Although the method in Reference [16] adopts shortcut connections, it can only deliver local features to nearby higher layers. Thus, many of the local features from the low layers are still lost in the final layer, which affects the precision of the keypoints.…”

“…We evaluated the performance of the keypoint detection network using the mean error [16]. Mean error is defined as the Euclidean distance between the predicted keypoints and the ground truths normalized by the length of the aircraft's fuselage.…”

“…Zhao et al [15] designed a keypoint detection model based on CNNs, and a keypoint matching method to recognize aircrafts. Furthermore, Zuo et al [16] built a template matching method using both aircrafts' keypoints and segmentation results, which provided more detailed information to improve recognition accuracy. Although these methods have achieved good performance on certain datasets, they cannot effectively tackle images without resolution, which severely limits their practicality.…”