Deep learning-based autofocus improvement metric for synthetic aperture sonar and synthetic aperture radar
A deep learning-based system and method for determining whether an autofocus algorithm has improved or degraded the image quality of a snippet derived from a synthetic aperture sonar system or a synthetic aperture radar system. The deep learning-based system and method embody a trained convolutional neural network that is configured to process a pair of snippets consisting of an unfocused snippet and a focused snippet. The focused snippet is produced by performing an autofocus algorithm on the unfocused snippet. The convolutional neural network processes the pair of snippets to determine if the autofocus algorithm has improved or degraded the image quality of the unfocused snippet. The convolutional neural network generates data that represents a confidence of improvement in image quality. The convolutional neural network leverages self-supervised learning and requires significantly fewer ground truth labels than a fully-supervised neural network.
1 . A method of establishing an image quality metric for an autofocused synthetic aperture image comprising the steps of:
receiving a synthetic aperture image and retaining said synthetic aperture image in processor memory;
autofocusing said received synthetic aperture image into an autofocused image and retaining said autofocused image in processor memory;
extracting features from said synthetic aperture image utilizing a feature extractor to provide an original feature vector;
extracting features from said autofocused image utilizing the same feature extractor to provide an autofocused feature vector;
concatenating the original feature vector and the autofocused feature vector to provide a concatenated feature vector;
conducting a batch normalization of the concatenated feature vector in a neural network resulting in a normalized feature vector;
reducing the number of elements in the normalized feature vector weighing the importance of different features utilizing a first linear neural network feature reduction function to provide a reduced feature vector;
applying a rectified linear unit activation function to the reduced feature vector to provide an adapted reduced feature vector; and
reducing the number of elements in the adapted reduced feature vector utilizing a second linear neural network feature reduction function to provide a single image quality metric indicative of the confidence that the autofocused image is of higher image quality than the image quality of the synthetic aperture image.
2 . The method of claim 1 wherein said step of reducing the number of elements in the adapted reduced feature vector further comprises reducing the number of elements in the adapted reduced feature vector to a single element and performing a sigmoid function on the single element to give the single image quality metric.
3 . The method of claim 2 wherein the synthetic aperture image is a synthetic aperture sonar image.
4 . The method of claim 2 wherein the synthetic aperture image is a synthetic aperture radar image.
5 . The method of claim 1 wherein the neural network is a ResNet-18 neural network with the three inputs input channels decreased to one high frequency input for receiving the concatenated feature vector.