Natural images are sparse signals in many transformation domains such as Fourier and Wavelet domains. With this observation, an image can be compressed by selecting a few dominant coefficients in the transformation domain for coding. This is the basic idea in compression standards such as JPEG, JPEG2000, and MPEG. A feature-specific imaging system directly collects the coefficients as measurements. We call the measurement feature. To collect one feature, the object is imaged onto a mask, and a detector is used to integrate and measure the illumination after the mask. Such a system is also called compressive imaging system some time. Compared with traditional camera imaging system, feature-specific imaging system has advantages such as less number of detectors, less hardware cost, smaller measurement noise, higher energy consumption efficiency, and less operation time. It has potential application areas including data reconstruction such as MRI and spectroscopy, object detection and tracking, and pattern recognition.
In this project, we consider different architectures for compressive imaging system implementation. We analyzed three architectures: the sequential, parallel, and pipeline architectures under noisy measurement condition. Object reconstruction is used as the application example. In a sequential architecture FSI system, there is only single detector to collect the measurements one after the other. Parallel architecture FSI system includes a detector array. The object is imaged by a lenslet onto a mask array. Each mask is controlled individually for each feat. Pipeline architecture system measures the features at the same time as in a sequential architecture system. Because the pipeline architecture system uses the object illumination most efficiently, it has the best reconstruction performance. Among all three architectures, the sequential architecture has the simplest hardware design and the worst performance. Different features such as discrete cosine transformation (DCT), Hadamard, principal component (PC), and random projected features are used for the analysis. Multiple non-linear reconstruction methods are also implemented to improve the system performance.
Distributed sensor network has attracted many attentions in recent years. We proposed to use non-imaging sensor or feature-specific (FS) sensor for distributed imaging network. In such sensor network, each sensor collects several features and sent them back to the base station. All features are processed at the base station for object reconstruction. We discussed sequential and parallel architectures for these sensors. The major advantage of using feature-specific sensor is that the networks lifetime with a fixed reconstruction performance requirement is improved.
In this project, we use FSI system for motion detection. Traditionally, to decide if there is motion in a field of view, a high resolution image is measured first. Then each image pixel is processed with detection algorithm. According to the number of moving pixels, a decision is made for detecting motion in the defined field of view. We directly collect very few features to accomplish the detection task. The adaptive Gaussian mixture model is used for modeling the feature measurements. Back-ground subtraction algorithm is used for detecting motion.
In this project, we design PCA and Hadamard features for each object using an adaptive procedure. We use a training set for the design procedure. Each feature is designed according to previous measurements. In each adaptation iteration, we select the training samples close to the testing sample in the measurement domain to form a new training set. Then the dominant features are selected for feature measurements in the next iteration. Using such a feature design procedure, the least number of features are used to accomplish a reconstruction performance requirement.