Search

In recent image classification approaches, a vision transformer (ViT) has shown an excellent performance beyond that of a convolutional neural network. A ViT achieves a high classification for natural images because it properly preserves the global image features. Conversely, a ViT still has many limitations in facial expression recognition (FER), which requires the detection of subtle changes in expression, because it can lose the local features of the image. Therefore, in this paper, we propose Squeeze ViT, a method for reducing the computational complexity by reducing the number of feature dimensions while increasing the FER performance by concurrently combining global and local features. To measure the FER performance of Squeeze ViT, experiments were conducted on lab-controlled FER datasets and a wild FER dataset. Through comparative experiments with previous state-of-the-art approaches, we proved that the proposed method achieves an excellent performance on both types of datasets. [ABSTRACT FROM AUTHOR]

The deep random forest (DRF) has recently gained new attention in deep learning because it has a high performance similar to that of a deep neural network (DNN) and does not rely on a backpropagation. However, it connects a large number of decision trees to multiple layers, thereby making analysis difficult. This paper proposes a new method for simplifying a black-box model of a DRF using a proposed rule elimination. For this, we consider quantifying the feature contributions and frequency of the fully trained DRF in the form of a decision rule set. The feature contributions provide a basis for determining how features affect the decision process in a rule set. Model simplification is achieved by eliminating unnecessary rules by measuring the feature contributions. Consequently, the simplified and transparent DRF has fewer parameters and rules than before. The proposed method was successfully applied to various DRF models and benchmark sensor datasets while maintaining a robust performance despite the elimination of a large number of rules. A comparison with state-of-the-art compressed DNNs also showed the proposed model simplification's higher parameter compression and memory efficiency with a similar classification accuracy. [ABSTRACT FROM AUTHOR]

As the need for wildfire detection increases, research on wildfire smoke detection combining low-cost cameras and deep learning technology is increasing. Camera-based wildfire smoke detection is inexpensive, allowing for a quick detection, and allows a smoke to be checked by the naked eye. However, because a surveillance system must rely only on visual characteristics, it often erroneously detects fog and clouds as smoke. In this study, a combination of a You-Only-Look-Once detector and a long short-term memory (LSTM) classifier is applied to improve the performance of wildfire smoke detection by reflecting on the spatial and temporal characteristics of wildfire smoke. However, because it is necessary to lighten the heavy LSTM model for real-time smoke detection, in this paper, we propose a new method for applying the teacher–student framework to deep LSTM. Through this method, a shallow student LSTM is designed to reduce the number of layers and cells constituting the LSTM model while maintaining the original deep LSTM performance. As the experimental results indicate, our proposed method achieves up to an 8.4-fold decrease in the number of parameters and a faster processing time than the teacher LSTM while maintaining a similar detection performance as deep LSTM using several state-of-the-art methods on a wildfire benchmark dataset. [ABSTRACT FROM AUTHOR]

Depth estimation is a crucial and fundamental problem in the computer vision field. Conventional methods re-construct scenes using feature points extracted from multiple images; however, these approaches require multiple images and thus are not easily implemented in various real-time applications. Moreover, the special equipment required by hardware-based approaches using 3D sensors is expensive. Therefore, software-based methods for estimating depth from a single image using machine learning or deep learning are emerging as new alternatives. In this paper, we propose an algorithm that generates a depth map in real time using a single image and an optimized lightweight efficient neural network (L-ENet) algorithm instead of physical equipment, such as an infrared sensor or multi-view camera. Because depth values have a continuous nature and can produce locally ambiguous results, pixel-wise prediction with ordinal depth range classification was applied in this study. In addition, in our method various convolution techniques are applied to extract a dense feature map, and the number of parameters is greatly reduced by reducing the network layer. By using the proposed L-ENet algorithm, an accurate depth map can be generated from a single image quickly and, in a comparison with the ground truth, we can produce depth values closer to those of the ground truth with small errors. Experiments confirmed that the proposed L-ENet can achieve a significantly improved estimation performance over the state-of-the-art algorithms in depth estimation based on a single image. [ABSTRACT FROM AUTHOR]

Human detection using visible surveillance sensors is an important and challenging work for intruder detection and safety management. The biggest barrier of real-time human detection is the computational time required for dense image scaling and scanning windows extracted from an entire image. This paper proposes fast human detection by selecting optimal levels of image scale using each level's adaptive region-of-interest (ROI). To estimate the image-scaling level, we generate a Hough windows map (HWM) and select a few optimal image scales based on the strength of the HWM and the divide-and-conquer algorithm. Furthermore, adaptive ROIs are arranged per image scale to provide a different search area. We employ a cascade random forests classifier to separate candidate windows into human and nonhuman classes. The proposed algorithm has been successfully applied to real-world surveillance video sequences, and its detection accuracy and computational speed show a better performance than those of other related methods.