Salient Region Guided Colour Image Restoration using Deep Learning with Adaptive Compressive Sensing

Wireless Visual Sensor Networks (WVSNs) have gained significant importance in the last few years and have emerged in several distinctive applications. The main aim is to design low power WVSN surveillance application using adaptive Compressive Sensing (CS) which is expected to overcome the WVSN resource constraints such as memory limitation, communication bandwidth and battery constraints. In this paper, an adaptive block CS technique is proposed and implemented to represent the high volume of captured images in a way for energy efficient wireless transmission and minimum storage. Furthermore, to achieve energy-efficient target detection and tracking with high detection reliability and robust tracking, to maximize the lifetime of sensor nodes as they can be left for months without any human interactions. Adaptive CS is expected to dynamically achieve higher compression rates depending on the sparsity nature of different datasets, while only compressing relative blocks in the image that contain the target to be tracked instead of compressing the whole image. Hence, saving power and increasing compression rates. Least mean square adaptive filter is used to predicts target's next location to investigate the effect of CS on the tracking performance. The tracking is achieved in both indoor and outdoor environments for single/multi targets. Results have shown that with adaptive block CS up to 20 % measurements of data are required to be transmitted while preserving the required performance for target detection and tracking.

Automatic detection of salient image regions is a useful tool with applications in intelligent camera control, virtual cinematography, video summarization and editing, evaluation of viewer preferences, and many others. This paper presents an effective method for detecting potentially salient foreground regions. Salient regions are identified by eigenvalue analysis of a graph Laplacian that is defined over the color similarity of image superpixels, under the assumption that the majority of pixels on the image boundary show non-salient background. In contrast to previous methods based on graph-cuts or graph partitioning, our method provides continuously-valued saliency estimates with complementary properties to recently proposed color contrast-based approaches. Moreover, exploiting discriminative properties of the Fiedler vector, we devise an SVM-based classifier that allows us to determine whether an image contains any salient objects at all, a problem that has been largely neglected in previous works. We also describe how the per-frame saliency detection can be extended to improve its spatiotemporal coherence when computed on video sequences. Extensive evaluation on several datasets demonstrates and validates the state-of-the-art performance of the proposed method.

Global feature integration based salient region detection

To solve problems involving sequential decision-making, Deep Reinforcement Learning (DRL) combines the advantages of Reinforcement Learning (RL) and Deep Learning (DL).These fusions are skilled at training agents for video games because they use neural networks to approximate nonlinear functions.The Deep Q Network (DQN), a key algorithm in this field, tends to overestimate Q-values despite its effectiveness.The Double Deep Q Network (Double DQN) was presented to address this.The Kera's framework in PyCharm is used in this study to examine the practical application and comparative analysis of DQN and Double DQN in the Flappy Bird game.To speed up its training, improvements were also made to the DQN model.The enhanced DQN performed better than the traditional DQN but less well than the Double DQN, according to the results.This study delves into the practical application and comparative analysis of DQN and Double DQN in the Flappy Bird game, utilizing the Kera's framework in PyCharm.Additionally, enhancements were made to the DQN model to expedite its training.Results affirmed that the enhanced DQN outperformed the conventional DQN but lagged the Double DQN.The training loss trajectory further substantiated Double DQN's superiority in mitigating overestimation issues.

Salient region detection is to uniformly locate interest regions or objects in an image. It is a hot topic in computer vision, and has a wide range of applications like object recognition and segmentation. Although considerable progress has been made, salient region detection remains a challenging issue. In this paper we present a simple yet effective salient region detection approach by integrating spatial and background priors. By considering spatial priors which include spatial distribution of color similarity and center-bias, the background regions and foreground regions are extracted preliminarily. Based on the extracted background we use the background prior to suppress the background regions more sufficiently. Experimental results on public benchmark databases show that the proposed approach can effectively locate salient object regions with well-defined boundaries and suppress background regions.

Selective Visual Attention Model (SVAM) plays an important role in region-based image retrieval. In this paper, a robust and accurate method for salient region detection is proposed which integrates SVAM and image segmentation. After that, the concept of salient region adjacency graphs (SRAGs) is introduced for image retrieval. The whole process consists of three levels. First in the pixel-level, the salient value of each pixel is calculated using an improved spatial-based attention model. Then in the region-level, the salient region detection method is presented. Furthermore, in the scene-level, salient region adjacency graphs (SRAGs) are introduced to represent the salient groups in the image, which take the salient regions as root nodes. Finally, the constructed SRAGs are used for image retrieval. Experiments show that the proposed method works well.KeywordsRoot NodeImage RetrievalSalient RegionImage Segmentation MethodAttention WindowThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

In this article, we tackle the saliency detection task from an interesting perspective: we focus both on salient regions (or foreground) detection and nonsalient regions (or background) detection instead of only the foreground and propose a novel complementarity-aware attention network. It is a unified framework with two branches, namely, positive attention module (PAM) and negative attention module (NAM), for the foreground and background detection, respectively. More specifically, the PAM exploits a position self-attention mechanism to enhance the discriminant ability of feature representation, which can detect most of the salient object regions. Meanwhile, the NAM is designed to detect the background regions, aiming to pop out the missing object parts and details in the prediction map produced by the PAM. By fusing these two attention modules together, NAM can provide complementary cues to assist PAM for precise object detection. Furthermore, in order to capture more multiscale contextual information, we introduce a bidirectional structure with multisupervision to the proposed complementarity-aware attention module for performance improvement. Experiments on five benchmark datasets show that the proposed framework achieves comparable results compared with the state-of-the-art saliency detection methods.

Compressive sensing and processing handles high-resolution delay-Doppler radar measurements using a low sampling rate, simple receiver design, and inexpensive processing when Nyquist rate sensing and processing becomes impractical. The benefits of compressive sensing and processing are, however, offset by an increase in estimation error that is introduced when processing compressed measurements versus measurements sampled at the Nyquist rate. In this work, an adaptive compressive sensing and processing method is proposed for the radar tracking problem. The adaptive scheme naturally incorporates information on target state that is readily available from a particle filter based tracker. The proposed method is shown to improve tracking performance over nonadaptive compressive sensing and processing, while maintaining low-sampling rates, a computationally inexpensive operation, and a simple receiver design.

In this paper, we propose a lightweight and adaptive compressed sensing method based on deep learning for edge devices (LACSLE) that generate and transmit sensor data using wireless connection. To transmit data efficiently, edge devices should change the compression rate adaptively according to data sparsity. However, a large amount of calculation is required to optimize the compression ratio due to the repeated processes of basis conversion and L1-norm minimization required for calculating the reconstruction error. To address this problem, LACSLE optimizes the compression ratio employed in compressed sensing in real time to maximize the data transmission efficiency using deep learning. In LACSLE, a deep learning model is trained in advance to map sensor data to an optimal compression ratio using either supervised learning (LACSLE-SL) or reinforcement learning (LACSLE-RL). The model is installed on an edge device and used to estimate an optimal compression ratio. The performance evaluation of LACSLE demonstrated that the proposed method allows to significantly reduce the computational delay of the optimal compression ratio estimation while reducing both the compression ratio and reconstruction error.

Finding most striking region in an image is known as salient region detection. This area is becoming an area of research in recent years due to its wide applicability in computer vision, robotics and data transmission. To find salient region various parameters like color, texture, location, semantics etc are used. In this paper, a method is proposed which uses the fact that image boundary is rarely touched by salient object and then poisson distribution is used to find the probability of each pixel being part of salient object. Thus, the saliency map is obtained by computing the absolute difference of pixel intensity with the mean of boundary pixels. No former training is essential for this approach. Experiments are performed to evaluate the performance of proposed approach on MSRA dataset and compared with 4 recent state-of-arts. Precision, recall and F-measure proves the consistency of proposed work.

Salient instance segmentation via subitizing and clustering

Object detection problems also expect fast and accurate location of target objects. But most approaches following convolutional neural networks (CNN) framework generate a large number of redundant candidate regions, wasting a large percent of computation time for candidate classification. Human visual system is able to quickly find the parts or features belonging to objects of interest. Hence, it is believed that by detecting salient regions associated with key classification features for specific objects can quickly provide rough location of objects to be detected. A feature mining method by adding an enforcement learning block on top of CNN framework is proposed. Experiment results showed that the proposed method is able to generate accurate task specific salient regions both for localization and classification. This study attempts to make object localization and classification in one shot based on salient region detection and puts forward a new idea for key feature mining and utilization.

Saliency region detection plays an important role in computer vision applications and related areas, such as human fixation, figure-ground separation, face detection, and image compression. Nevertheless, state-of-the-art methods can moderately detect the saliency regions of a psychological pattern dataset. Therefore, this paper proposes a spatial-frequency redistribution (SFR) method to improve the efficiency of the detection of the saliency regions. The proposed SFR method consists of two main procedures: (i) adaptive cosine image construction-and-implantation and (ii) saliency region detection using complete multiple-object-implanted images. The former procedure constructs an adaptive cosine image by using a cosine function based on an object structure and then individually implants it into each object detectable. This stage provides the first significant property, spatial redistribution, to the object implanted. The adaptive cosine image redistributes the objects for controllability. Then, the latter procedure transforms a complete multiple object-implanted image into the frequency domain. At this point, the second significant property, frequency redistribution, provides the simple technique for identifying and separating the target object from the unwanted objects and background. In this paper, these properties are referred to as spatial-frequency redistribution. This method was realized as a computer program, and then such program was tested with a psychological pattern dataset and a Thai text dataset. The experimental results, when compared with state-of-the-art methods, show that the proposed SFR method can achieve the clear detection of the attention region in the psychological pattern dataset. Moreover, the proposed method can achieve a F-value of 80.28% with a recall and precision of 76.32% and 85.23%, respectively, for Thai text localisation in the Thai text dataset.

A salient region is the most distinctive part of the image that captures human's attention. Saliency detection is a fundamental characteristic of the human visual system. Finding computational models which are able to detect salient regions is a challenging task for image processing and computer vision applications. Salient regions of various sizes can be detected from different scales. Therefore, selecting the best scales is an important issue. In this paper, an efficient multi-scale method to find salient regions is proposed. In order to include more features in evaluating saliency of a pixel, feature maps are generated using components of both the RGB and YUV color spaces. These features are combined into quaternions. Detecting salient regions of different sizes is addressed by utilizing a scale space analysis. Salient regions are detected by convolving the image amplitude spectrum with a low-pass Gaussian kernel of multiple scales. To incorporate more meaningful information, more than one scale is considered based on entropy criterion. The final saliency map is generated by normalizing the weighted saliency maps of these scales. Experiments are conducted on a dataset of natural images to evaluate the performance of the proposed method. Results show that the proposed method provides larger values of area under receiver operating characteristics curve, precision, recall and F-measure, in comparison to some of the state-of-the-art methods.

This paper devises a novel adaptive framework for the energy-aware acquisition of spectrally sparse signals. The adaptive quantized compressive sensing (CS) techniques, beyond-complementary metal–oxide–semiconductor (CMOS) hardware architecture, and corresponding algorithms which utilize them have been designed concomitantly to minimize the overall cost of signal acquisition. First, a spin-based adaptive intermittent quantizer (AIQ) is developed to facilitate the realization of the adaptive sampling technique. Next, a framework for smart and adaptive determination of the sampling rate and quantization resolution based on the instantaneous signal and hardware constraints is introduced. Finally, signal reconstruction algorithms which process the quantized CS samples are investigated. Simulation results indicate that an AIQ architecture using a spin-based quantizer incurs only 20.98- $\mu \text{W}$ power dissipation on average using 22-nm technology for 1–8 bits uniform output. Furthermore, in order to provide 8-bit quantization resolution, 85.302- $\mu \text{W}$ maximum power dissipation is attained. Our results indicate that the proposed AIQ design provides up to 6.18-mW power savings on average compared to other adaptive rate and resolution CMOS-based CS analog-to-digital converter designs. In addition, the mean square error values achieved by the simulation results confirm efficient reconstruction of the signal based on the proposed approach.