Active Learning Based on Crowdsourced Data

The paper proposes a crowdsourcing-based approach for annotated data acquisition and means to support Active Learning training approach. In the proposed solution, aimed at data engineers, the knowledge of the crowd serves as an oracle that is able to judge whether the given sample is informative or not. The proposed solution reduces the amount of work needed to annotate large sets of data. Furthermore, it allows a perpetual increase in the trained network quality by the inclusion of new samples, gathered after network deployment. The paper also discusses means of limiting network training times, especially in the post-deployment stage, where the size of the training set can increase dramatically. This is done by the introduction of the fourth set composed of samples gather during network actual usage.

A Review on Content Based Image Retrieval System Features derived by Deep Learning Models

In a Content Based Image Retrieval (CBIR) System, the task is to retrieve similar images from a large database given a query image. The usual procedure is to extract some useful features from the query image, and retrieve images which have similar set of features. For this purpose, a suitable similarity measure is chosen, and images with high similarity scores are retrieved. Naturally the choice of these features play a very important role in the success of this system, and high level features are required to reduce the “semantic gap”. In this paper, we propose to use features derived from pre-trained network models from a deep- learning convolution network trained for a large image classification problem. This approach appears to produce vastly superior results for a variety of databases, and it outperforms many contemporary CBIR systems. We analyse the retrieval time of the method, and also propose a pre-clustering of the database based on the above-mentioned features which yields comparable results in a much shorter time in most of the cases. Keywords Content Based Image Retrieval Feature Selection Deep Learning Pre-trained Network Models Pre-clustering

Recognition and calculation of objects in images using YOLOv3 architecture

Program that searches for five types of fruits in the images of fruit trees, classifies them and counts their quantity is presented. Its creation took into account the requirement to be able to work both in the background and in real time and to identify the desired objects at a sufficiently high speed. The program should also be able to learn from available computers (including laptops) and within a reasonable time. In carrying out this task, the possibilities of several existing approaches to the recognition and identification of visual objects based on the use of convolutional neural networks were analyzed. Among the considered network archi-tectures were R-CNN, Fast R-CNN, Faster R-CNN, SSD, YOLO and some modifications based on them. Based on the analysis of the peculiarities of their work, the YOLO architecture was used to perform the task, which allows the analy-sis of visual objects in real time with high speed and reliability. The software product was implemented by modifying the YOLOv3 architecture implemented in TensorFlow 2.1. Object recognition in this architecture is performed using a trained Darknet-53 network, the parameters of which are freely available. The modification of the network was to replace its original classification layer. The training of the network modified in this way was carried out on the basis of Transfer learning technology using the Agrilfruit Dataset. There was also a study of the peculiarities of the learning process of the network under the use of different types of gradient descent (stochastic and with the value of the batch 4 and 8), as a result of which the optimal version of the trained network weights was selected for further use. Tests of the modified and trained network have shown that the system based on it with high reliability distin-guishes objects of the corresponding classes of different sizes in the image (even with their significant masking) and counts their number. The ability of the program to distinguish and count the number of individual fruits in the analyzed image can be used to visually assess the yield of fruit trees

ERDBNet: Enhanced Residual Dense Block Net --A New Net to Rich ESRGAN Image Details

Super resolution is applied in many digital image fields. In many cases, only a set of low-resolution images can be obtained, but the image needs a higher resolution, and then SR needs to be applied. SR technology has undergone years of development. Among them, SRGAN is the key work to introduce GAN into the SR field, which can truly restore a large number of details on the basis of low-pixel pictures. ESRGAN is a further improvement on SRGAN. By removing the BN layer in SRGAN, the effect of artifacts in SRGAN is eliminated. However, there is still a problem that the restoration of information on small and medium scales is not accurate enough. The proposed ERDBNet improve the model on the basis of ESRGAN, and use the ERDB block to replace the original RRDB block. The new structure uses a three-layer dense block to replace the original dense block, and a residual structure of the starting point is added to each dense block. The pre-trained network can reach a PSNR of 30.425 after 200k iterations, and the minimum floating PSNR is only 30.213. Compared with the original structure, it is more stable and performs better in the detail recovery of many low-pixel images.

Improved visual inertial odometry based on deep learning

The traditional visual inertial odometry according to the manually designed rules extracts key points. However, the manually designed extraction rules are easy to be affected and have poor robustness in the scene of illumination and perspective change, resulting in the decline of positioning accuracy. Deep learning methods show strong robustness in key point extraction. In order to improve the positioning accuracy of visual inertial odometer in the scene of illumination and perspective change, deep learning is introduced into the visual inertial odometer system for key point detection. The encoder part of MagicPoint network is improved by depthwise separable convolution, and then the network is trained by self-supervised method; A visual inertial odometer system based on deep learning is compose by using the trained network to replace the traditional key points detection algorithm on the basis of VINS. The key point detection network is tested on HPatches dataset, and the odometer positioning effect is evaluated on EUROC dataset. The results show that the improved visual inertial odometer based on deep learning can reduce the positioning error by more than 5% without affecting the real-time performance.

Near-Field-Based 5G Sub-6 GHz Array Antenna Diagnosis Using Transfer Learning

In this paper, we propose a method for near-field-based 5G sub 6-GHz array antenna diagnosis using transfer learning. A classification network was implemented for normal/abnormal operation of the array antenna and the failure of a specific port. Furthermore, a regression network that could predict the amplitude and phase of the excitation signal of the array antenna was employed. Additionally, to accelerate the array antenna diagnosis, several near-field lines were sampled and reflected in the regression network. The proposed method was verified by measuring a fabricated 5G sub-6 GHz band 4×4 array antenna in various scenarios using a divider and coaxial cables. The tests showed that the trained network accurately diagnosed 29 of 30 measurement results.

DEEP CONVOLUTIONAL NEURAL NETWORK FOR SEGMENTING CT IMAGES OF FIBER REINFORCEMENTS

The greatest challenge in creating digital material twins and FE mesh from μCT images of composite reinforcements is the lack of a robust and versatile tool for training μCT images. Here, we have used deep convolutional neural networks (DCNN) for segmenting μCT images of a multi-layer plain-weave fiber reinforcement. A set of raw 2D image slices extracted from the gray-scale volume of a single-layer reinforcement was used to train a DCNN using manually annotated images. The trained network was tested against the manually segmented ground truth images and it performed exceptionally well with a global accuracy of more than 96%. The trained DCNN was then used to segment unseen images from a multilayer stack of the fabric with good accuracy. The work presented here provides a robust and efficient framework of segmenting CT scan images of fiber reinforcements for generating digital material twins and FE mesh of fiber reinforcements.

Best Practices for the Deployment of Edge Inference: The Conclusions to Start Designing

The number of Artificial Intelligence (AI) and Machine Learning (ML) designs is rapidly increasing and certain concerns are raised on how to start an AI design for edge systems, what are the steps to follow and what are the critical pieces towards the most optimal performance. The complete development flow undergoes two distinct phases; training and inference. During training, all the weights are calculated through optimization and back propagation of the network. The training phase is executed with the use of 32-bit floating point arithmetic as this is the convenient format for GPU platforms. The inference phase on the other hand, uses a trained network with new data. The sensitive optimization and back propagation phases are removed and forward propagation is only used. A much lower bit-width and fixed point arithmetic is used aiming a good result with reduced footprint and power consumption. This study follows the survey based process and it is aimed to provide answers such as to clarify all AI edge hardware design aspects from the concept to the final implementation and evaluation. The technology as frameworks and procedures are presented to the order of execution for a complete design cycle with guaranteed success.