Soft computing

About: Soft computing is a research topic. Over the lifetime, 6710 publications have been published within this topic receiving 118508 citations.

Type-2 Fuzzy Logic for Improving Training Data and Response Integration in Modular Neural Networks for Image Recognition

Abstract: The combination of Soft Computing techniques allows the improvement of intelligent systems with different hybrid approaches. In this work we consider two parts of a Modular Neural Network for image recognition, where a Type-2 Fuzzy Inference System (FIS 2) makes a great difference. The first FIS 2 is used for feature extraction in training data, and the second one to find the ideal parameters for the integration method of the modular neural network. Once again Fuzzy Logic is shown to be a tool that can help improve the results of a neural system, when facilitating the representation of the human perception.

Updated discussions on ‘Hybrid multiple criteria decision-making methods: a review of applications for sustainability issues’

Abstract: A recent review discussed a variety of hybrid multiple criteria decision-making (HMCDM) methods on the subject of sustainability issues. Some soft computing techniques, such as the fuzzy set, have ...

A hybrid bio-inspired learning algorithm for image segmentation using multilevel thresholding

Abstract: In the field of image analysis, segmentation is one of the most important preprocessing steps. One way to achieve segmentation is the use of threshold selection, where each pixel that belongs to a determined class, based on the mutual visual characteristics, is labeled according to the selected threshold. In this work, a combination of two pioneer methods, namely Otsu and Kapur, are investigated to solve the threshold selection problem. Optimum parameters of these objective functions are calculated using Bacterial Foraging (BF) optimization algorithm, for its accuracy, and Harmony Search (HS), for its speed. However, the biggest problem of soft computing family algorithms is catching into a local optimum. To resolve this critical issue, we investigate the power of Learning Automata (LA) which works as a controller to make switching between these two optimization methods. LA is a heuristic method which can solve complex optimization problems with interesting results in parameter estimation. Despite other techniques commonly seek through the parameter map, LA explores in the probability space, providing appropriate convergence properties and robustness. The proposed method is tested on benchmark images and shows fast convergence avoiding the typical sensitivity to initial conditions such as the Expectation-Maximization (EM) algorithm or the complex, and time-consuming computations which are commonly found in gradient methods. Experimental results demonstrate the algorithm's ability to perform automatic multi-threshold selection and show interesting advantages as it is compared to other algorithms solving the same task.

Deep Convolutional Networks for Construction Object Detection Under Different Visual Conditions

Abstract: Sensing and reality capture devices are widely used in construction sites. Among different technologies, vision-based sensors are by far the most common and ubiquitous. A large volume of images and videos is collected from construction projects every day to track work progress, measure productivity, litigate claims, and monitor safety compliance. Manual interpretation of such colossal amounts of data, however, is non-trivial, error-prone, and resource-intensive. This has motivated new research on soft computing methods that utilize high-power data processing, computer vision, and deep learning (DL) in the form of convolutional neural networks (CNN). A fundamental step toward machine-driven interpretation of construction site scenery is to accurately identify objects of interest for a particular problem. The accuracy requirement, however, may offset the computational speed of the candidate method. While light-weight DL algorithms (e.g., Mask R-CNN) can perform visual recognition with relatively high accuracy, they suffer from low processing efficacy which hinders their use in real-time decision-making. One of the most promising DL algorithms that balance speed and accuracy is YOLO (you-only-look-once). This paper investigates YOLO-based CNN models in fast detection of construction objects. First, a large-scale image dataset, named Pictor-v2, is created which contains about 3,500 images and approximately 11,500 instances of common construction site objects (e.g., building, equipment, worker). To assess the agility of object detection, transfer learning is used to train two variations of this model, namely YOLO-v2 and YOLO-v3, and test them on different data combinations (crowdsourced, web-mined, or both). Results indicate that performance is higher if the model is trained on both crowdsourced and web-mined images. Additionally, YOLO-v3 outperforms YOLO-v2 by focusing on smaller, harder-to-detect objects. The best performing YOLO-v3 model has a 78.2% mAP when tested on crowdsourced data. Sensitivity analysis of the output shows that the model’s strong suit is in detecting larger objects in less crowded and well-lit spaces. The proposed methodology can be also extended to predict the relative distance of the detected objects with reliable accuracy. Findings of this work lay the foundation for further research on technology-assistive systems to augment human capacities in quickly and reliably interpreting visual data in complex environments.