Fixed Window Size Research Articles

SwinTExCo: Exemplar-based video colorization using Swin Transformer

Video colorization represents a compelling domain within the field of Computer Vision. The traditional approach in this field relies on Convolutional Neural Networks (CNNs) to extract features from each video frame and employs a recurrent network to learn information between video frames. While demonstrating considerable success in colorization, most traditional CNNs suffer from a limited receptive field size, capturing local information within a fixed-sized window. Consequently, they struggle to directly grasp long-range dependencies or pixel relationships that span large image or video frame areas. To address this limitation, recent advancements in the field have leveraged Vision Transformer (ViT) and their variants to enhance performance. This article introduces Swin Transformer Exemplar-based Video Colorization (SwinTExCo), an end-to-end model for the video colorization process that incorporates the Swin Transformer architecture as the backbone. The experimental results demonstrate that our proposed method outperforms many other state-of-the-art methods in both quantitative and qualitative metrics. The achievements of this research have significant implications for the domain of documentary and history video restoration, contributing to the broader goal of preserving cultural heritage and facilitating a deeper understanding of historical events through enhanced audiovisual materials.

Multi-resolution short-time Fourier transform providing deep features for 3D CNN to classify rolling bearing fault vibration signals

The time-frequency domain features of vibration signals provide valuable information for deep learning-based rolling bearing fault diagnosis methods, where fault signal classification aiding in the identification of nominal fault types during diagnosis. The Short-Time Fourier Transform (STFT) is a widely used time-frequency transformation method, and its window length is the key parameter that determines the trade-off between time and frequency resolution. The primary motivation of this study is to address the limitation in traditional STFT-based 2D CNN methods: the inability to adapt the window length to different types of signals. To achieve accurate classification of bearing fault types, this study proposes a method based on three-dimensional convolutional neural networks (3D CNNs) to deeply explore the time-frequency domain information of one-dimensional vibration signals from faulty bearings. This method first applies STFT with multiple window sizes to perform multi-resolution time-frequency transformations on the time-domain vibration signals, yielding three-dimensional data. Subsequently, a classifier is trained based on the proposed 3D CNN. Experimental results on public datasets show that, without any sophisticated techniques, the proposed method achieves an average classification accuracy of 99.2% for six types of bearing faults using a relatively simple CNN structure. Compared to 1D CNN and 2D CNN methods that use fixed window sizes for STFT, the proposed method significantly enhances classification performance. Furthermore, it demonstrates robust classification results even on small-scaled bearing datasets.

Graph feature fusion driven by deep autoencoder for advanced hyperspectral image unmixing

In this paper, we propose a pioneering approach for blind Hyperspectral Image (HSI) unmixing named Multi-Features Graph Deep Fusion Learning Networks for HSI Unmixing (MF-GDL). Our method leverages the power of multi-feature graph deep fusion learning networks. MF-GDL integrates spectral HSI graph features with complementary spatial graph features, overcoming the fixed window size limitation observed in prior deep learning approaches in capturing spatial relationships effectively. A key contribution lies in the introduction of multi-view graph learning, a novel aspect in HSI unmixing. By incorporating graph similarity-based spatial information, we address an important gap in the existing literature, marking the first attempt, to our knowledge, to tackle this challenge using a graph deep learning method. Furthermore, we introduce auxiliary information based on Extended Morphological Profiles (EMPs) to enrich spatial context beyond high-level spatial relationships between neighboring pixels. This multi-view fusion learning not only improves accuracy in material abundance estimation but also demonstrates the significance of considering multiple perspectives in HSI unmixing. We validate the efficacy of MF-GDL on four real datasets: Jasper Ridge, Samson, Urban, and Apex showcasing its robust performance and highlighting the substantial contribution of the multi-view aspect in advancing HSI unmixing techniques.

SDIPPWV: A novel hybrid prediction model based on stepwise decomposition-integration-prediction avoids future information leakage to predict precipitable water vapor from GNSS observations

Water vapor is an important meteorological parameter. Accurate prediction of water vapor content can be used to provide important reference information for heavy rainfall forecast and artificial precipitation operation. The current water vapor hybrid prediction model has the problem of future data leakage, and the error is accumulated by reconstructing the subsequence after prediction. Therefore, this paper proposes a stepwise decomposition-integration-prediction precipitable water vapor mechanism, named SDIPPWV, which can effectively solve the above problems. Firstly, High-precision precipitable water vapor (PWV) sequence is retrieved from Global Navigation Satellite System (GNSS) observation files. Then stepwise decomposition process uses a fixed-size window to segment the PWV sequence and Seasonal-Trend decomposition based on Loess (STL) to decompose the sequences within the window. Next, the features of the three sub-sequences are integrated to construct the feature space. Finally the prediction of PWV is obtained using 1D Convolutional Neural Network-Bidirectional Long Short Term Memory (1D CNN-BiLSTM). The model performance is verified using observation data from eight GNSS stations. The performance of the PWV prediction model proposed in this paper is effectively improved compared with the single prediction models and other hybrid models. The mean root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE) and coefficient of determination (R2) of the eight stations are 0.2146 mm, 0.1132 mm, 1.29 %, and 0.9998, respectively. The results show that the model proposed in this paper improves the prediction accuracy of water vapor content while solving the data leakage problem.

EA-GAT: Event aware graph attention network on cyber-physical systems

Anomaly detection with high accuracy, recall, and low error rate is critical for the safe and uninterrupted operation of cyber-physical systems. However, detecting anomalies in multimodal time series with different modalities obtained from cyber-physical systems is challenging. Although deep learning methods show very good results in anomaly detection, they fail to detect anomalies according to the requirements of cyber-physical systems. In the use of graph-based methods, data loss occurs during the conversion of time series into graphs. The fixed window size used to transform time series into graphs causes a loss of spatio-temporal correlations. In this study, we propose an Event Aware Graph Attention Network (EA-GAT), which can detect anomalies by event-based cyber-physical system analysis. EA-GAT detects and tracks the sensors in cyber-physical systems and the correlations between them. The system analyzes and models the relationship between the components during the marked periods as a graph. Anomalies in the system are found through the created graph models. Experiments show that the EA-GAT technique is more effective than other deep learning methods on SWaT, WADI, MSL datasets used in various studies. The event-based dynamic approach is significantly superior to the fixed-size sliding window technique, which uses the same learning structure. In addition, anomaly analysis is used to identify the attack target and the affected components. At the same time, with the slip subsequence module, the data is divided into subgroups and processed simultaneously.

All thresholds barred: direct estimation of call density in bioacoustic data

Passive acoustic monitoring (PAM) studies generate thousands of hours of audio, which may be used to monitor specific animal populations, conduct broad biodiversity surveys, detect threats such as poachers, and more. Machine learning classifiers for species identification are increasingly being used to process the vast amount of audio generated by bioacoustic surveys, expediting analysis and increasing the utility of PAM as a management tool. In common practice, a threshold is applied to classifier output scores, and scores above the threshold are aggregated into a detection count. The choice of threshold produces biased counts of vocalizations, which are subject to false positive/negative rates that may vary across subsets of the dataset. In this work, we advocate for directly estimating call density: The proportion of detection windows containing the target vocalization, regardless of classifier score. We propose a validation scheme for estimating call density in a body of data and obtain, through Bayesian reasoning, probability distributions of confidence scores for both the positive and negative classes. We use these distributions to predict site-level call densities, which may be subject to distribution shifts (when the defining characteristics of the data distribution change). These methods may be applied to the outputs of any binary detection classifier operating on fixed-size audio input windows. We test our proposed methods on a real-world study of Hawaiian birds and provide simulation results leveraging existing fully annotated datasets, demonstrating robustness to variations in call density and classifier model quality.

PWLT: Pyramid Window-based Lightweight Transformer for image classification

Recently, vision Transformers (ViTs) have achieved remarkable progress for image classification. As the computational cost of self-attention adopted in ViTs is quadratic with respect to the number of input tokens, some window-based ViTs have been proposed to solve this issue. However, these methods limit the computation of self-attention into spatial-constrained local windows, losing capability to encode image-based global interactions. Additionally, using fixed-size window always suffers the limitation of single-scale representation that is unsuitable for object recognition with variable scales. To address these problems, this paper describes a Pyramid Window-based Lightweight Transformer, namely PWLT, for image classification. Specifically, to address the need for multi-scale information, we employ windows of different sizes to encode objects with varying scales. To restore the relationships between different windows and explore global context, we introduce a dual self-attention scheme that utilizes local-to-global attention to reestablish these relationships. The extensive experiments on ImageNet-1K and CIFAR100 datasets demonstrate the effectiveness of our PWLT for image classification.

PMFFNet: A hybrid network based on feature pyramid for ovarian tumor segmentation.

Ovarian cancer is a highly lethal malignancy in the field of oncology. Generally speaking, the segmentation of ovarian medical images is a necessary prerequisite for the diagnosis and treatment planning. Therefore, accurately segmenting ovarian tumors is of utmost importance. In this work, we propose a hybrid network called PMFFNet to improve the segmentation accuracy of ovarian tumors. The PMFFNet utilizes an encoder-decoder architecture. Specifically, the encoder incorporates the ViTAEv2 model to extract inter-layer multi-scale features from the feature pyramid. To address the limitation of fixed window size that hinders sufficient interaction of information, we introduce Varied-Size Window Attention (VSA) to the ViTAEv2 model to capture rich contextual information. Additionally, recognizing the significance of multi-scale features, we introduce the Multi-scale Feature Fusion Block (MFB) module. The MFB module enhances the network's capacity to learn intricate features by capturing both local and multi-scale information, thereby enabling more precise segmentation of ovarian tumors. Finally, in conjunction with our designed decoder, our model achieves outstanding performance on the MMOTU dataset. The results are highly promising, with the model achieving scores of 97.24%, 91.15%, and 87.25% in mACC, mIoU, and mDice metrics, respectively. When compared to several Unet-based and advanced models, our approach demonstrates the best segmentation performance.

Expanding Sparse Radar Depth Based on Joint Bilateral Filter for Radar-Guided Monocular Depth Estimation.

Radar data can provide additional depth information for monocular depth estimation. It provides a cost-effective solution and is robust in various weather conditions, particularly when compared with lidar. Given the sparse and limited vertical field of view of radar signals, existing methods employ either a vertical extension of radar points or the training of a preprocessing neural network to extend sparse radar points under lidar supervision. In this work, we present a novel radar expansion technique inspired by the joint bilateral filter, tailored for radar-guided monocular depth estimation. Our approach is motivated by the synergy of spatial and range kernels within the joint bilateral filter. Unlike traditional methods that assign a weighted average of nearby pixels to the current pixel, we expand sparse radar points by calculating a confidence score based on the values of spatial and range kernels. Additionally, we propose the use of a range-aware window size for radar expansion instead of a fixed window size in the image plane. Our proposed method effectively increases the number of radar points from an average of 39 points in a raw radar frame to an average of 100 K points. Notably, the expanded radar exhibits fewer intrinsic errors when compared with raw radar and previous methodologies. To validate our approach, we assess our proposed depth estimation model on the nuScenes dataset. Comparative evaluations with existing radar-guided depth estimation models demonstrate its state-of-the-art performance.

Performance evaluation of individual tree detection and segmentation algorithms using ALS data in Chir Pine (Pinus roxburghii) forest

The application of individual tree detection algorithms for assessing forest inventories and aiding decision-making in forestry has been a subject of research for more than two decades. Nevertheless, there is a notable research gap in the development of robust algorithms capable of automatically detecting trees of different species, ages, and varied crown sizes in dense forest environments. In this study, we conducted a comprehensive evaluation of six different individual tree detection (ITD) algorithms using airborne LiDAR data in Chir Pine (Pinus roxburghii) forests. This research represents one of the pioneering efforts in applying ITD algorithms to Chir Pine forests using ALS data. We categorized ITD routines into two groups: those reliant on local maxima as treetops and initial seeds for crown delineation, and those that do not require explicit treetop identification. To assess accuracy, we developed a special Individual Tree Matching (ITM) algorithm, enabling the matching of LiDAR-detected trees with 284 reference trees measured in the field. Our analysis involved various combinations of filtering fixed window sizes and adaptive window sizes applied to the point cloud, unsmoothed, and smoothed canopy height model (CHM). Our results highlighted the effectiveness of the 3 × 3m fixed window size method on unsmoothed CHM, achieving an overall F-score of 0.65 and a tree detection rate of 86%. Additionally, the Dalponte 2016 method proved superior for crown segmentation using identified treetops, consistently measuring mean crown radii within 0.5 m of reference field trees. Among the methods not relying on treetops, the adaptive mean shift algorithm (AMS3D) delivered strong performance, boasting an overall F-score of 0.67 and mean crown radii within 0.1 m. Our study revealed a high correlation between LiDAR-detected tree heights and field-measured tree heights across all evaluated methods. Overall, our findings underscore the potential of ITD algorithms in enhancing forest attribute measurement accuracy and facilitating climate-responsive forest management strategies.

CADM + : Confusion-Based Learning Framework With Drift Detection and Adaptation for Real-Time Safety Assessment.

Real-time safety assessment (RTSA) of dynamic systems holds substantial implications across diverse fields, including industrial and electronic applications. However, the complexity and rapid flow nature of data streams, coupled with the expensive label cost and pose significant challenges. To address these issues, a novel confusion-based learning framework, termed confusion-and-detection method plus (CADM + ), is proposed in this article. When drift occurs, the model is updated with uncertain samples, which may cause confusion between existing and new concepts, resulting in performance differences. The cosine similarity is used to measure the degree of such conceptual confusion in the model. Furthermore, the change of standard deviation within a fixed-size cosine similarity window is introduced as an indicator for drift detection. Theoretical demonstrations show the asymptotic increase of cosine similarity. In addition, the approximate independence of the change in standard deviation with the number of trained samples is indicated. Finally, the extreme value theory (EVT) is applied to determine the threshold of judging drifts. Several experiments are conducted to verify its effectiveness. Experimental results prove that the proposed framework is more suitable for RTSA tasks compared with state-of-the-art algorithms. The source code is available at https://github.com/THUFDD/CADM-plus.

On the influence of the supposed global deep fault on the strong earthquakes in Indonesia

According to the NEIC global catalog, out of 14 earthquakes with a magnitude of MS ≥7.0 registered in the world in the first six months of 2023, two belong to the powerful multi-earthquake of Turkey 06.02.2023 (Mw: 7.8 и 7.7), and the remaining 12 occurred in the Indonesian region. In the paper the geodynamic situation in the preparation of a series of Indonesian events, which began with two earthquakes: 08 and 09.01.2023 (MS : 7.0 и 7.6, Н=29 и 105 km), is investigated by means of a geoinformation system by means of a geoformation system GIS-ENDDB. It is established that all 12 earthquakes are spatially confined to the sub-latitudinal “Indonesian” seismolineament with a length of 9500 km, detected by the method of the great circle of the Earth. The dynamics of changes in the geophysical environment properties of the 950-kilometer zone from the axis of the lineament was estimated by the value of the paired correlation of magnitude MS and creepex Cr0_ISC. Two methods were used to calculate the coefficient of paired correlation of these parameters in a sliding time window: 1) with a fixed window size and 2) with one edge fixed at zero time (corresponding to the moment of the strongest Indonesian event on 09.01.2023). Both methods confirm the fact of monotonous consolidation of the environment along the lineament (evenly along its entire length and depth) 33 days before and 31 days after the strongest Indonesian shock: at the interval of reverse correlation (|KKOR|≥0.7) of magnitude and creepex. The same pattern was previously described by us for the consolidated part of the earth’s crust, where stronger earthquakes have a smaller creepex (characterizing the greater contribution of brittle destruction). Thus, the fact of the formation of an organized state of the environment (linear rigid structure) within the limits of the assumed deep fault, affecting the subsequent seismicity of the region in the form of a spatially timed semi-annual series of the strongest earthquakes with MS ≥7.0 is confirmed.

Two-constraint-free dual-domain optimised random phase-only hologram

To overcome the main limitations of the optimised random phase (ORAP) non-iterative hologram generation method, which include the support constraints (position, size, and shape) and the fixed window size of the space domain, we propose a two-constraint-free dual-domain optimised random phase (TCF-DD-ORAP) method for phase-only hologram generation. Firstly, the dual-domain optimised random phase (DD-ORAP) is generated in the dual domain (the space or frequency domain), both of which can be used to non-iteratively generate holograms for different target images. Next, the support constraints of the target image are relaxed in the space domain (position, size, and shape), and the generated DD-ORAP can then be used to generate holograms for the target images with arbitrary support in the image space. Finally, a crop operation is applied to target images with different window sizes, and the TCF-DD-ORAP corresponding to the target window size can be cropped for CGH generation. The TCF-DD-ORAP method proposed in this paper overcomes the dual limitations of fixed support and window size that arise in the original ORAP method of CGH generation, and can therefore significantly reduce the computational cost and enhance the flexibility of CGH generation. Numerical and optical experiments validate the effectiveness and superiority of our proposed TCF-DD-ORAP method.

A physical exertion inspired multi-task learning framework for detecting out-of-breath speech

Physical exertion is a stress condition that affects how we normally produce speech. It alters both the temporal and spectral pattern of speech characteristics. Therefore, speech utterances can be used as a cost-effective telehealth solution to detect whether a person is under physical exertion. This paper deals with the detection of shortness of breath (or out-of-breath) condition from speech under physical load by using an multi-task learning (MTL) framework. The primary classification targets are the neutral and the out-of-breath classes. Naturally these are binary targets and do not reflect the actual extent of exertion. This leads to the creation of a novel auxiliary target learning in collaboration with a pre-trained expert system. The targets here indicate the level of exertion under physical load. The MTL framework for both the convolutional neural network (CNN) and CNN-long short-term memory (CLSTM) network performs nearly 1.5% (F1-score) better than the single task learning (STL) framework. In addition to that, out-of-breath speech has more influence on the lower frequency spectrum. Warped spectrograms are given as input to the networks, enabling the deep networks to focus on lower spectral regions. The non-linear frequency warping is achieved by Mel-scale transformation and constant-Q-transform (CQT). CQT, being less dependent on fixed-window size, shows at least 6.57% (F1-score) improvement over Mel-based inputs. The MTL framework, combined with the warped spectrum, performs better in classifying out-of-breath speech from neutral.

Counting Activities Using Weakly Labeled Raw Acceleration Data: A Variable-Length Sequence Approach with Deep Learning to Maintain Event Duration Flexibility.

This paper presents a novel approach for counting hand-performed activities using deep learning and inertial measurement units (IMUs). The particular challenge in this task is finding the correct window size for capturing activities with different durations. Traditionally, fixed window sizes have been used, which occasionally result in incorrectly represented activities. To address this limitation, we propose segmenting the time series data into variable-length sequences using ragged tensors to store and process the data. Additionally, our approach utilizes weakly labeled data to simplify the annotation process and reduce the time to prepare annotated data for machine learning algorithms. Thus, the model receives only partial information about the performed activity. Therefore, we propose an LSTM-based architecture, which takes into account both the ragged tensors and the weak labels. To the best of our knowledge, no prior studies attempted counting utilizing variable-size IMU acceleration data with relatively low computational requirements using the number of completed repetitions of hand-performed activities as a label. Hence, we present the data segmentation method we employed and the model architecture that we implemented to show the effectiveness of our approach. Our results are evaluated using the Skoda public dataset for Human activity recognition (HAR) and demonstrate a repetition error of ±1 even in the most challenging cases. The findings of this study have applications and can be beneficial for various fields, including healthcare, sports and fitness, human-computer interaction, robotics, and the manufacturing industry.

People Recognition via Tongue Print Using Deep and Machine Learning

The tongue is a unique organ that is well protected inside the mouth and not affected by external factors; it is also difficult to forge. Several biometric systems are widely used for authentication and recognition, such as fingerprints, faces, iris, sound, retina, etc. Traditional biometrics represent a challenge and an obstacle as they can be falsified, duplicates can be made (e.g., iris, face, fingers, signature), or they are expensive and rarely used (e.g., DNA). The increased security measures called for modern biometrics that is more secure, less expensive, and cannot be falsified. As a result, the goal of this paper is to create a system for distinguishing people based on their tongue prints. It will contribute to solving many forensic issues and increasing electronic security because it has features suitable for identification and biometrically distinguishing between people. In this paper, the tongue is located based on the fixed window size method. After tongue localization (ROI), feature extraction using the VGG-16 model, and a classification system that uses both transfer learning and machine learning as VGG-16, XGBoost, KNN, and RF classifiers, extracted features are then trained for personal identification. The dataset consisted of 1085 tongue images of 138 people with a test ratio of 20%, and the results achieved an accuracy of 92%. The process of distinguishing people through tongue prints has proven to be effective and accurate.

Online Parameter Identification for Lithium-Ion Batteries: An Adaptive Moving Window Size Design Methodology for Least Square Fitting

Online state-of-charge estimations are critical to lithium-ion batteries in electric vehicles for safe and reliable operations. To ensure accurate online state estimations, an online battery parameter identification algorithm, such as the least square fitting approach, along with a shifting moving window, was developed and implemented, while the decision of the moving window size for general purpose is a technical issue not yet been well-addressed. This paper proposes an adaptive moving window size design methodology that adjusts the moving window size considering three important factors: load profile, state-of-charge to open-circuit voltage profile, and condition of data. The simulation results show that the proposed methodology helps to identify the parameters of the battery electric circuit model with high accuracy, where the results are compatible to the results using fine-tuned fixed window sizes. Also, the proposed methodology is general to driving profiles from moderate to severe without extra algorithm redesign. In the three standard driving profile cases, the parameter identification error is reduced by 59.15%, 69.02%, and 75.14%, respectively, compared to the worst case in experiments, which are compatible with the results after fine-tuning the fixed window sizes. With accurately identified parameters, the modeling error is minimized, and three typical linear SOC estimation algorithms thus render accurate SOC estimation, rendering an average of 13.4%, 64.43% 57.48% reduction from the worst case in experiments, respectively.

Financial Time Series Forecasting: A Data Stream Mining-Based System

Data stream mining (DSM) represents a promising process to forecast financial time series exchange rate. Financial historical data generate several types of cyclical patterns that evolve, grow, decrease, and end up dying. Within historical data, we can notice long-term, seasonal, and irregular trends. All these changes make traditional static machine learning models not relevant to those study cases. The statistically unstable evolution of financial market behavior yields a progressive deterioration in any trained static model. Those models do not provide the required characteristics to evolve continuously and sustain good forecasting performance as the data distribution changes. Online learning without DSM mechanisms can also miss sudden or quick changes. In this paper, we propose a possible DSM methodology, trying to cope with that instability by implementing an incremental and adaptive strategy. The proposed algorithm includes the online Stochastic Gradient Descent algorithm (SGD), whose weights are optimized using the Particle Swarm Optimization Metaheuristic (PSO) to identify repetitive chart patterns in the FOREX historical data by forecasting the EUR/USD pair’s future values. The data trend change is detected using a statistical technique that studies if the received time series instances are stationary or not. Therefore, the sliding window size is minimized as changes are detected and maximized as the distribution becomes more stable. Results, though preliminary, show that the model prediction is better using flexible sliding windows that adapt according to the detected distribution changes using stationarity compared to learning using a fixed window size that does not incorporate any techniques for detecting and responding to pattern shifts.

An Adaptive Identification Method for Potential Landslide Hazards Based on Multisource Data

The effectiveness of landslide disaster prevention depends largely on the quality of early identification of potential hazards, and how to comprehensively, deeply, and accurately identify such hazards has become a major difficulty in landslide disaster management. Existing deep learning methods for potential landslide hazard identification often use fixed-size window modeling and ignore the different window sizes required by landslides of different scales. To address this problem, we propose an adaptive identification method for potential landslide hazards based on multisource data. Taking Yongping County, China, as the study area, we create a multisource factor dataset based on the landslide disaster background in terms of topography, geology, human activities, hydrology, and vegetation as the sample for the identification model after processing. Moreover, we combine differential interferometric synthetic aperture radar (D-InSAR) and multitemporal InSAR (MT-InSAR) to process the surface deformation of the study area, and we measure the deformation richness based on the average of the pixel deformation difference within the current window of a pixel point in the image. Therefore, convolutional neural networks (CNNs) with different window sizes are adaptively selected. The results show that the precision of adaptive identification of potential landslide hazards in the study area is 85.30%, the recall is 83.03%, and the F1 score is 84.15%. The recognition rate for potential hazards reaches 80%, which is better than the fixed-window modeling result and proves the effectiveness of the proposed method. This method can help to improve intelligent identification systems for potential landslide hazards, and also contribute to the identification of other potential geological hazards, such as mudslides and collapses.

A Deep Learning-Based Semantic Segmentation Model Using MCNN and Attention Layer for Human Activity Recognition.

With the development of wearable devices such as smartwatches, several studies have been conducted on the recognition of various human activities. Various types of data are used, e.g., acceleration data collected using an inertial measurement unit sensor. Most scholars segmented the entire timeseries data with a fixed window size before performing recognition. However, this approach has limitations in performance because the execution time of the human activity is usually unknown. Therefore, there have been many attempts to solve this problem through the method of activity recognition by sliding the classification window along the time axis. In this study, we propose a method for classifying all frames rather than a window-based recognition method. For implementation, features extracted using multiple convolutional neural networks with different kernel sizes were fused and used. In addition, similar to the convolutional block attention module, an attention layer to each channel and spatial level is applied to improve the model recognition performance. To verify the performance of the proposed model and prove the effectiveness of the proposed method on human activity recognition, evaluation experiments were performed. For comparison, models using various basic deep learning modules and models, in which all frames were classified for recognizing a specific wave in electrocardiography data were applied. As a result, the proposed model reported the best F1-score (over 0.9) for all kinds of target activities compared to other deep learning-based recognition models. Further, for the improvement verification of the proposed CEF method, the proposed method was compared with three types of SW method. As a result, the proposed method reported the 0.154 higher F1-score than SW. In the case of the designed model, the F1-score was higher as much as 0.184.