Although unsupervised representation learning (RL) can tackle the performance deterioration caused by limited labeled data in synthetic aperture radar (SAR) object classification, the neglected discriminative detailed information and the ignored distinctive characteristics of SAR images can lead to performance degradation. In this paper, an unsupervised multi-scale convolution auto-encoder (MSCAE) was proposed which can simultaneously obtain the global features and local characteristics of targets with its U-shaped architecture and pyramid pooling modules (PPMs). The compact depth-wise separable convolution and the deconvolution counterpart were devised to decrease the trainable parameters. The PPM and the multi-scale feature learning scheme were designed to learn multi-scale features. Prior knowledge of SAR speckle was also embedded in the model. The reconstruction loss of the MSCAE was measured by the structural similarity index metric (SSIM) of the reconstructed data and the images filtered by the improved Lee sigma filter. A speckle suppression restriction was also added in the objective function to guarantee that the speckle suppression procedure would take place in the feature learning stage. Experimental results with the MSTAR dataset under the standard operating condition and several extended operating conditions demonstrated the effectiveness of the proposed model in SAR object classification tasks.

Abstract.With the development of image processing and storage technology, rapid classification and annotation of huge volumes of digital images have been attracted much attention. However, the complex and ambiguous relationship between images and concept classes poses significant challenges on building effective annotation models. Structured machine learning methods have been studied to tackle the problem of complex relationship between concept classes for prediction, which have been proved effective for image understanding tasks. We proposed a novel image annotation model based on structured machine learning, by introducing a learned kernel function in the sample space, aiming at capturing the underlying distribution of concept classes of the training data set. The model is evaluated on two benchmark data sets and the results show that the model is promising compared to current state-of-the-art methods.

No abstract