Whereas single class classification has been a highly active topic in optical remote sensing, much less effort has been given to the multi-label classification framework, where pixels are associated with more than one labels, an approach closer to the reality than single-label classification. Given the complexity of this problem, identifying representative features extracted from raw images is of paramount importance. In this work, we investigate feature learning as a feature extraction process in order to identify the underlying explanatory patterns hidden in low-level satellite data for the purpose of multi-label classification. Sparse auto-encoders composed of a single hidden layer, as well as stacked in a greedy layer-wise fashion formulate the core concept of our approach. The results suggest that learning such sparse and abstract representations of the features can aid in both remote sensing and multi-label problems. The results presented in the paper correspond to a novel real dataset of annotated spectral imagery naturally leading to the multi-label formulation.
Acquisition of high dimensional Hyperspectral Imaging (HSI) data using limited dimensionality imaging sensors has led to restricted capabilities designs that hinder the proliferation of HSI. To overcome this limitation, novel HSI architectures strive to minimize the strict requirements of HSI by introducing computation into the acquisition process. A framework that allows the integration of acquisition with computation is the recently proposed framework of Compressed Sensing (CS). In this work, we propose a novel HSI architecture that exploits the sampling and recovery capabilities of CS to achieve a dramatic reduction in HSI acquisition requirements. In the proposed architecture, signals from multiple spectral bands are multiplexed before getting recorded by the imaging sensor. Reconstruction of the full hyperspectral cube is achieved by exploiting a dictionary of elementary spectral profiles in a unified minimization framework. Simulation results suggest that high quality recovery is possible from a single or a small number of multiplexed frames.
Compressed Sensing (CS) is a novel mathematical framework that has revolutionized modern signal and image acquisition architectures ranging from one-pixel cameras, to range imaging and medical ultrasound imaging. According to CS, a sparse signal, or a signal that can be sparsely represented in an appropriate collection of elementary examples, can be recovered from a small number of random linear measurements. However, real life systems may introduce non-linearities in the encoding in order to achieve a particular goal. Quantization of the acquired measurements is an example of such a non-linearity introduced in order to reduce storage and communications requirements. In this work, we consider the case of scalar quantization of CS measurements and propose a novel recovery mechanism that enforces the constraints associated with the quantization processes during recovery. The proposed recovery mechanism, termed Quantized Orthogonal Matching Pursuit (Q-OMP) is based on a modification of the OMP greedy sparsity seeking algorithm where the process of quantization is explicit considered during decoding. Simulation results on the recovery of images acquired by a CS approach reveal that the modified framework is able to achieve significantly higher reconstruction performance compared to its naive counterpart under a wide range of sampling rates and sensing parameters, at a minimum cost in computational complexity.
Active range imaging (RI) systems utilize actively controlled light sources emitting laser pulses that are subsequently recorded by an imaging system and used for depth profile estimation. Classical RI systems are limited by their need for a large number of frames required to obtain high resolution depth information. In this work, we propose an RI approach motivated by the recently proposed compressed sensing framework to dramatically reduce the number of necessary frames. Compressed gated range sensing employs a random gating mechanism along with state-of-the-art reconstruction algorithms for the estimation of the timing of the reflected pulses and the inference of distances. In addition to efficiency, the proposed scheme is also able to identify multiple reflected pulses that can be introduced by semi-transparent elements in the scene such as clouds, smoke, and foliage. Simulations under highly realistic conditions demonstrate that the proposed architecture is capable of accurately recovering the depth profile of a scene from as few as 10 frames at 100 depth bins resolution, even under very challenging conditions. The results further indicate that the proposed architecture is able to extract multiple reflected pulses with a minimal increase in the number of frames, in situations where state-of-the-art methods fail to accurately estimate the correct depth signals.
Active Range Imaging (ARI) has recently sparked an enthusiastic interest due to the numerous applications that can benefit from the high quality depth maps that ARI systems offer. One of the most successful ARI techniques employs Time-of-Flight (ToF) cameras which emit and subsequently record laser pulses in order to estimate the distance between the camera and objects in a scene. A limitation of this type of ARI is the requirement for a large number of frames that have to be captured in order to generate high resolution depth maps. In this work, we introduce Compressed Gated Range Sensing (CGRS), a novel approach for ToF-based ARI that utilizes the recently proposed framework of Compressed Sensing (CS) to dramatically reduce the number of necessary frames. The CGRS technique employs a random gating function along with state-of-the-art reconstruction in order to estimate the timing of a returning laser pulse and infer the depth map. To validate our method, software simulations were carried out using a realistic system model. Simulated results suggest that low error reconstruction of a depth map is possible using approximately 20% of the frames that traditional ToF cameras require, while 30% sampling rates can achieve very high fidelity reconstruction.
Range Imaging (RI) has sparked an enthusiastic interest recently due to the numerous applications that can benefit from the presence 3D data. One of the most successful techniques for RI employs Time-of-Flight (ToF) cameras which emit and subsequently record laser pulses in order to estimate the distance between the camera and an object. A limitation of this class of RI is the requirement for a large number of frames that have to be captured in order to generate high resolution depth maps. In this work, we propose a novel approach for ToF based RI that utilizes the recently proposed framework of Compressed Sensing to dramatically reduce the number of necessary frames. Our technique employs a random gating function along with state-of-the-art minimization techniques in order to estimate the location of a returning laser pulse and infer the distance. To validate the theoretical motivation, software simulations were carried out. Our simulated results have shown that reconstruction of a depth map is possible from as low as 10% of the frames that traditional ToF cameras require with minimum reconstruction error while 20% sampling rates can achieve almost perfect reconstruction in low resolution regimes. Our experimental results have also shown that the proposed method is robust to various types of noise and applicable to realistic signal models.