Automatically detecting dangerous concealed objects in the millimeter-wave images is important for imaging-aided security systems. In this paper, we proposed a normalized accumulation map-based training mechanism for concealed object detection network. The proposed normalized accumulation map, calculated as the average of binary masks representing the object location for each image, can reveal the positions of frequently-appeared concealed objects, which offers different weights for different locations when computing confidence loss. Experiments on a millimeter-wave security image dataset demonstrate the effectiveness of our proposed normalized accumulation map-based training mechanism. By introducing our training mechanism to YOLO-v2, the object detection network can achieve a 4.43% performance improvement in mean average precision.

Phase unwrapping is a critical step in synthetic aperture radar interferometry (InSAR) data processing chains. In almost all phase unwrapping methods, estimating the phase gradient according to the phase continuity assumption (PGE-PCA) is an essential step. The phase continuity assumption is not always satisfied due to the presence of noise and abrupt terrain changes; therefore, it is difficult to get the correct phase gradient. In this paper, we propose a robust least squares phase unwrapping method that works via a phase gradient estimation network based on the encoder–decoder architecture (PGENet) for InSAR. In this method, from a large number of wrapped phase images with topography features and different levels of noise, the deep convolutional neural network can learn global phase features and the phase gradient between adjacent pixels, so a more accurate and robust phase gradient can be predicted than that obtained by PGE-PCA. To get the phase unwrapping result, we use the traditional least squares solver to minimize the difference between the gradient obtained by PGENet and the gradient of the unwrapped phase. Experiments on simulated and real InSAR data demonstrated that the proposed method outperforms the other five well-established phase unwrapping methods and is robust to noise.

In the problem of geospatial object detection, the quality and amount of reference boxes significantly impact the detection performance and prediction speed of object detection networks. Nowadays, most of the popular detection methods adopt the anchor mechanism to generate reference boxes. This paper proposed an anchor-free and sliding-window-free deconvolutional region proposal network and constructed a two-stage deconvolutional object detection network. Instead of using an anchor mechanism, we proposed to use a deconvolutional neural network followed by a connected region generation module to generate reference boxes. The comparison experiments and quantitative analysis with NWPU VHR-10 dataset demonstrate that DeRPN can vastly reduce the number of reference boxes and improve the precision of the reference box coordinates. The experiments also suggest that our proposed two-stage object detection network can not only obtain the nearly state-of-the-art detection results but also achieve the prediction speed close to that of the one-stage detection network.

The success of deep learning in synthetic aperture radar (SAR) automatic target recognition (ATR) relies on a large number of labeled samples; however, there are often wrong (noisy) labels in a large-scale dataset. In this article, we propose a loss curve-fitting-based method, which can identify the noisy labels and train the classification network effectively. We propose to model label noise by unsupervised clustering via fitting loss curve to identify whether the sample’s label is clean or noisy. Then, we train the network using augmented samples with clean labels to correct noisy labels further. The experiments on the moving and stationary target acquisition and recognition (MSTAR) dataset prove that our proposed method can deal with the situation when training a network with different ratios of noisy labels and correct noisy labels effectively. When the noise ratio is small (40%) in the training dataset, our method can correct 97.9% of noisy labels and train the classification network with 98.8% classification accuracy. While the noise ratio is large (80%), our method can correct 78.1% of noisy labels and train the classification network with 79.6% classification accuracy.

Because of the three-dimensional (3D) imaging scene’s sparsity, compressed sensing (CS) algorithms can be used for linear array synthetic aperture radar (LASAR) 3D sparse imaging. CS algorithms usually achieve high-quality sparse imaging at the expense of computational efficiency. To solve this problem, a fast Bayesian compressed sensing algorithm via relevance vector machine (FBCS–RVM) is proposed in this paper. The proposed method calculates the maximum marginal likelihood function under the framework of the RVM to obtain the optimal hyper-parameters; the scattering units corresponding to the non-zero optimal hyper-parameters are extracted as the target-areas in the imaging scene. Then, based on the target-areas, we simplify the measurement matrix and conduct sparse imaging. In addition, under low signal to noise ratio (SNR), low sampling rate, or high sparsity, the target-areas cannot always be extracted accurately, which probably contain several elements whose scattering coefficients are too small and closer to 0 compared to other elements. Those elements probably make the diagonal matrix singular and irreversible; the scattering coefficients cannot be estimated correctly. To solve this problem, the inverse matrix of the singular matrix is replaced with the generalized inverse matrix obtained by the truncated singular value decomposition (TSVD) algorithm to estimate the scattering coefficients correctly. Based on the rank of the singular matrix, those elements with small scattering coefficients are extracted and eliminated to obtain more accurate target-areas. Both simulation and experimental results show that the proposed method can improve the computational efficiency and imaging quality of LASAR 3D imaging compared with the state-of-the-art CS-based methods.

Compressed sensing (CS) algorithms are used for linear array synthetic aperture radar (LASAR) three-dimensional (3D) imaging. However, it is difficult to obtain imaging results with both high computational efficiency and promising imaging quality. Because of the high-dimensional matrix-operations, the computational complexity of several CS algorithms is huge such as the iterative adaptive approach (IAA), bayesian compressed sensing (BCS), and sparsity bayesian recovery via iterative minimum (SBRIM) algorithm. Besides, the greedy pursuit algorithms such as the orthogonal matching pursuit (OMP) algorithm cannot acquire ideal imaging results on account of the preset sparsity of the imaging scene. To solve the problem, we present a fast sparse recovery algorithm via resolution approximation (FSRARA) in this paper. Firstly, the whole imaging scene is divided into 3D scattering units with large spacing, and SBRIM algorithm is used to obtain its low-resolution imaging results quickly. Secondly, the low-resolution imaging results are conducted image segmentation by the fuzzy c-means (FCM) clustering algorithm to extract the possible targets' areas coarsely. Then we re-divide the imaging scene by higher imaging resolution and extract the possible targets' areas according to the coarsely possible targets' areas. FSRARA achieves improved computational efficiency with low-dimensional matrix-operations on the possible targets' areas instead of the high-dimensional one on the whole imaging scene. Meanwhile, FSRARA performs better in suppressing the false targets and sidelobe interference and improves the imaging quality than the SBRIM algorithm. Simulation and experimental results prove that FSRARA improves the computational efficiency by hundreds of times at most than SBRIM algorithm and its computational efficiency is higher than smoothed $L_{0} $ norm (SL0), IAA, and BCS algorithm. Besides, FSRARA improves the imaging quality compared with OMP, IAA, SL0, BCS, and SBRIM algorithms.