We consider the compression of multidimensional signals on the aircraft board. We describe the data of such signals as a hypercube, which is "rotated" in a special way. To compress this hypercube, we use a h...
详细信息
AI-powered edge devices currently lack the ability to adapt their embedded inference models to the ever-changing envi ronment. To tackle this issue, Continual Learning (CL) strategies aim at incrementally improving th...
详细信息
ISBN:
(纸本)9781728180991
AI-powered edge devices currently lack the ability to adapt their embedded inference models to the ever-changing envi ronment. To tackle this issue, Continual Learning (CL) strategies aim at incrementally improving the decision capabilities based on newly acquired data. In this work, after quantifying memory and computational requirements of CL algorithms, we define a novel HW/SW extreme-edge platform featuring a low power RISC-v octa-core cluster tailored for on-demand incremental learning over locally sensed data. The presented multi-core HW/SW architecture achieves a peak performance of 2.21 and 1.70 MAC/cycle, respectively, when running forward and backward steps of the gradient descent. We report the trade-off between memory footprint, latency, and accuracy for learning a new class with Latent Replay CL when targeting an image classification task on the CORe50 dataset. For a CL setting that retrains all the layers, taking 5h to learn a new class and achieving up to 77.3% of precision, a more efficient solution retrains only part of the network, reaching an accuracy of 72.5% with a memory requirement of 300 MB and a computation latency of 1.5 hours. On the other side, retraining only the last layer results in the fastest (867 ms) and less memory hungry (20 MB) solution but scoring 58% on the CORe50 dataset. Thanks to the parallelism of the low-power cluster engine, our HW/SW platform results 25× faster than typical MCU device, on which CL is still impractical, and demonstrates an 11× gain in terms of energy consumption with respect to mobile-class solutions.
There are many intelligent systems and tools which uses highly efficient processing models to identify different anomalies with high accuracy. The anomaly detection is of high importance and mostly will come as an abs...
详细信息
ISBN:
(数字)9781728196565
ISBN:
(纸本)9781728196572
There are many intelligent systems and tools which uses highly efficient processing models to identify different anomalies with high accuracy. The anomaly detection is of high importance and mostly will come as an absolute requirement at high risk environments and situations. The amount of processing involved in quick decision taking systems bare high deployment costs which restricts the anomaly detection only to a selected few who are capable of building such resource centered systems. Modern world uses drones and other video feeds in order to find and keep track of any anomalous events around a specific area. But most such detection requires absolute manual attention as well as processing power to keep up with real time detection and recognition. The proposed research solution aims to automate this process and includes a two-step anomaly detection system which gives a quicker anomaly detection in an average processing unit time with an advanced recognition model with up to 90% accuracy. The deep learning model (vGG 16) together with alert system and comparison techniques on videos leads into unsupervised anomaly detection of a landscape. The system generates alerts and recognizes anomalies on the alerted video frames. The proposed solution can also be used by any source and does not require high capacity of capability system to get the optimal output. Moreover, the solution brings a simple yet sophisticated technique to address modern anomaly detection and quick alerting system.
We consider minimizing the sum of three convex functions, where the first one F is smooth, the second one is nonsmooth and proximable and the third one is the composition of a nonsmooth proximable function with a line...
详细信息
Recently, surveillance cameras are ubiquitous for both real-time monitoring and recording important moments. Temporarily seamless surveillance using multiple cameras requires increasing amount of human efforts and eno...
ISBN:
(数字)9781728151861
ISBN:
(纸本)9781728151878
Recently, surveillance cameras are ubiquitous for both real-time monitoring and recording important moments. Temporarily seamless surveillance using multiple cameras requires increasing amount of human efforts and enormous size of storage. The use of dynamic cameras further requires advanced computer vision algorithms, and is another challenge for intelligent visual surveillance. To solve those problems, we present an enhanced metadata extraction method for robust object search and a person re-identification. More specifically, the proposed method accurately extracts an object region using a modified DeepLab version 3, and then extracts metadata including representative color, size, aspect ratio, and moving trajectory of the object. The proposed metadata extraction method can be applied to a wide range of surveillance systems such as search for missing children in a large public space and crowd monitoring system.
The article is devoted to the definition of such groups in social networks. The object of the study was selected data social network vk. Text data was collected, processed and analyzed. To solve the problem of obtaini...
详细信息
Document Classification is the problem of assigning documents into predefined category making it easier to manage and sort documents. There are several algorithms using which the problem of classification can be solve...
详细信息
The problem of unauthorized access to various content is an important task for the present. It becomes even more acute when using the existing toolkit for the processing and transmission of encrypted images with the f...
详细信息
The problem of unauthorized access to various content is an important task for the present. It becomes even more acute when using the existing toolkit for the processing and transmission of encrypted images with the fluctuation intensity function. The authors developed a modification of the RSA algorithm to use it in relation to the above -mentioned images. It is proposed a conceptual view for the joint use of quaternary fractional -linear transformations with elements of the basic RSA algorithm. The application of this mathematical apparatus in the basic RSA algorithm avoids contours of image objects in an encrypted sample. In addition, this combination provides additional stability to the basic RSA algorithm for unauthorized decryption. The simulation of the method was carried out in two described by the authors algorithms: using one line and four lines of the image matrix, for grayscale and color images. The high efficiency of the developed method for avoiding contours of objects on encrypted images has been confirmed. In both cases, contours after applying encryption procedures do not appear. The reverse procedure allows to get an image without visible distortion. 2019 The Authors. Published by Elsevier B.v.
Optimization of the radio thermal radiation noise processing in multichannel spatially distributed radiometer systems is performed. The algorithms consideration is applicable in the aperture synthesis systems. The max...
详细信息
Light field imaging technology has been attracting increasing interest because it enables capturing enriched visual information and expands the processing capabilities of traditional 2D imaging systems. Dense multivie...
详细信息
ISBN:
(纸本)9781538613115
Light field imaging technology has been attracting increasing interest because it enables capturing enriched visual information and expands the processing capabilities of traditional 2D imaging systems. Dense multiview, accurate depth maps and multiple focus planes are examples of different types of visual information enabled by light fields. This technology is also emerging in medical imaging research, like dermatology, allowing to find new features and improve classification algorithms, namely those based on machine learning approaches. This paper presents a contribution for the research community, in the form of a publicly available light field image dataset of skin lesions (named SKINL2 v1.0). This dataset contains 250 light fields, captured with a focused plenoptic camera and classified into eight clinical categories, according to the type of lesion. Each light field is comprised of 81 different views of the same lesion. The database also includes the dermatoscopic image of each lesion. A representative subset of 17 central view images of the light fields is further characterised in terms of spatial information (SI), colourfulness (CF) and compressibility. This dataset has high potential for advancing medical imaging research and development of new classification algorithms based on light fields, as well as in clinically-oriented dermatology studies.
暂无评论