Video Dataset Research Articles

Towards Visual-Prompt Temporal Answer Grounding in Instructional Video.

Temporal answer grounding in instructional video (TAGV) is a new task naturally derived from temporal sentence grounding in general video (TSGV). Given an untrimmed instructional video and a text question, this task aims at locating the frame span from the video that can semantically answer the question, i.e., visual answer. Existing methods tend to solve the TAGV problem with a visual span-based predictor, taking visual information to predict the start and end frames in the video. However, due to the weak correlations between the semantic features of the textual question and visual answer, current methods using the visual span-based predictor do not work well in the TAGV task. In this paper, we propose a visual-prompt text span localization (VPTSL) method, which introduces the timestamped subtitles for a text span-based predictor. Specifically, the visual prompt is a learnable feature embedding, which brings visual knowledge to the pre-trained language model. Meanwhile, the text span-based predictor learns joint semantic representations from the input text question, video subtitles, and visual prompt feature with the pre-trained language model. Thus, the TAGV is reformulated as the task of the visual-prompt subtitle span localization for the visual answer. Extensive experiments on five instructional video datasets, namely MedVidQA, TutorialVQA, VehicleVQA, CrossTalk and Coin, show that the proposed method outperforms several state-of-the-art (SOTA) methods by a large margin in terms of mIoU score, which demonstrates the effectiveness of the proposed visual prompt and text span-based predictor.

RoadSitu: Leveraging Road Video Frame Extraction and Three-Stage Transformers for Situation Recognition

Situation recognition is an crucial problem in scene understanding, activity understanding, and action reasoning as it provides a structured representation of the main activity depicted in the image.Semantic role labeling is crucial to situation recognition, which is challenging because a single action can have multiple meanings and purposes depending on its context. Understanding images beyond the highlighted actions requires inferences about the context of the scene, the objects, and their role in the captured event. Recently, situation recognition (SR) has been introduced, which jointly derives a collection of the action (activity), meaning-role, and noun (entities) pairs in the form of moving images. To label these frames as action frames, we must assign nouns (entities) to the role based on the content of the observed image. One of the main challenges is managing the complex dependencies between the assigned roles (nouns) and the predicted action, as the correct role assignment often depends on the accuracy of the action prediction. We introduce, RoadSitu, a road situation recognition that involves generating a structured summary of what is happening in a road scenario using an action and the semantic roles played by agents from a video frame. The action can describe a diverse set of situations, and the same agent can play various roles depending on the situation depicted in the video frame. Therefore, a situation recognition model needs to understand the context of each video frame and the visual-linguistic meaning of the semantic roles of that particular frame. One of the main challenges in this work is the complex task of annotating video frames with semantic roles and handling the structured dependencies between the assigned roles (nouns) and the predicted action (activity). Additionally, the sparsity of meaningful semantic information within road scenarios poses further difficulties. To overcome these challenges, we introduce a novel approach where action recognition and noun estimation work together interactively to form structured summaries of each situation. In experiments using a road video dataset obtained from a South Korean company, RoadSitu achieved significant improvements across various performance metrics, with a Top-1 verb accuracy of 43.46%, Top-5 verb accuracy of 72.48%, and value accuracy of 34.21%, outperforming baseline models such as GSRTR and JSL by 2.4% and 3.86% in Top-1 verb accuracy, respectively. These results demonstrate the effectiveness of our model in handling complex road scenarios.

Behavior Tracking and Analyses of Group-Housed Pigs Based on Improved ByteTrack

Daily behavioral analysis of group-housed pigs provides critical insights into early warning systems for pig health issues and animal welfare in smart pig farming. In this study, our main objective was to develop an automated method for monitoring and analyzing the behavior of group-reared pigs to detect health problems and improve animal welfare promptly. We have developed the method named Pig-ByteTrack. Our approach addresses target detection, Multi-Object Tracking (MOT), and behavioral time computation for each pig. The YOLOX-X detection model is employed for pig detection and behavior recognition, followed by Pig-ByteTrack for tracking behavioral information. In 1 min videos, the Pig-ByteTrack algorithm achieved Higher Order Tracking Accuracy (HOTA) of 72.9%, Multi-Object Tracking Accuracy (MOTA) of 91.7%, identification F1 Score (IDF1) of 89.0%, and ID switches (IDs) of 41. Compared with ByteTrack and TransTrack, the Pig-ByteTrack achieved significant improvements in HOTA, IDF1, MOTA, and IDs. In 10 min videos, the Pig-ByteTrack achieved the results with 59.3% of HOTA, 89.6% of MOTA, 53.0% of IDF1, and 198 of IDs, respectively. Experiments on video datasets demonstrate the method’s efficacy in behavior recognition and tracking, offering technical support for health and welfare monitoring of pig herds.

Learned scalable video coding for humans and machines

AbstractVideo coding has traditionally been developed to support services such as video streaming, videoconferencing, digital TV, and so on. The main intent was to enable human viewing of the encoded content. However, with the advances in deep neural networks (DNNs), encoded video is increasingly being used for automatic video analytics performed by machines. In applications such as automatic traffic monitoring, analytics such as vehicle detection, tracking and counting, would run continuously, while human viewing could be required occasionally to review potential incidents. To support such applications, a new paradigm for video coding is needed that will facilitate efficient representation and compression of video for both machine and human use in a scalable manner. In this manuscript, we introduce an end-to-end learnable video codec that supports a machine vision task in its base layer, while its enhancement layer, together with the base layer, supports input reconstruction for human viewing. The proposed system is constructed based on the concept of conditional coding to achieve better compression gains. Comprehensive experimental evaluations conducted on four standard video datasets demonstrate that our framework outperforms both state-of-the-art learned and conventional video codecs in its base layer, while maintaining comparable performance on the human vision task in its enhancement layer.

Spatter detection and tracking in high-speed video observations of laser powder bed fusion

Purpose In laser powder bed fusion (L-PBF) additive manufacturing, spatter particles are ejected from the melt pool and can be detrimental to material performance and powder recycling. Quantifying spatter generation with respect to processing conditions is a step toward mitigating spatter and better understanding the phenomenon. This paper reveals process insights of spatter phenomena by automatically annotating spatter particles in high-speed video observations using machine learning. Design/methodology/approach A high-speed camera was used to observe the L-PBF process while varying laser power, laser scan speed and scan strategy on a constant geometry on an EOSM290 using Ti-6Al-4V powder. Two separate convolutional neural networks were trained to segment and track spatter particles in captured high-speed videos for spatter characterization. Findings Spatter generation and ejection angle significantly differ between keyhole and conduction mode melting. High laser powers lead to large ejections at the beginning of scan lines. Slow and fast build rates produce more spatter than moderate build rates at constant energy density. Scan strategies with more scan vectors lead to more spatter. The presence of powder significantly increases the amount of spatter generated during the process. Originality/value With the ability to automatically annotate a large volume of high-speed video data sets with high accuracy, an experimental design of observed parameter changes reveals quantitively stark changes in spatter morphology that can aid process development to mitigate spatter occurrence and impacts on material performance.

Abstract 4120990: Classification of Echocardiography Videos Using TimeSformer for Detecting Incipient Heart Failure in Asymptomatic Patients with Normal Ejection Fraction and Patients with Heart Failure

Background: Recent advancements have seen deep learning models help differentiate echocardiography images of patients with heart failure with preserved ejection fraction (HFpEF) from normal controls. Our aim was to develop a model capable of detecting early signs of heart failure in asymptomatic patients with a normal ejection fraction. Methods: We employed the TimeSformer, a video transformer model that classifies video data using a novel attention-based mechanism. This self-attention mechanism that diverges from traditional convolutional neural networks (CNNs). It focuses on relevant parts of the video across both space and time, split into spatial attention, which processes each frame individually, and temporal attention, which integrates information across different frames. The training and validation of the TimeSformer model were conducted on the same dataset of echocardiography videos from 50 normal controls and 50 patients diagnosed with HFpEF, employing 5-fold cross-validation to ensure robust performance evaluation. Results: The TimeSformer model effectively identified HFpEF in patients, as all diagnosed with HFpEF were flagged as abnormal. The echocardiography assessment, along with NT pro BNP levels, supported the diagnoses, with patients showing NT pro BNP levels of 1016±32 pg/ml. Conversely, 46 out of 50 normal controls were correctly identified, with their NT pro BNP levels averaging 52±12 pg/ml. These controls remained asymptomatic. Conclusions: The TimeSformer model demonstrated capability in identifying subtle deviations indicative of incipient heart failure in videos of normal controls, despite normal NT pro BNP levels. This suggests potential for early detection of heart failure in asymptomatic individuals. Further longitudinal studies are necessary to determine whether individuals identified as abnormal develop heart failure symptoms subsequently.

A Video-based Lifting Action Recognition Using Rank-Altered Kinematic Feature Pairs

Low-back musculoskeletal disorders (MSDs) are the primary work-related injuries among manual material handling (MMH) workers, who are frequently exposed to repetitive lifting. To prevent low-back MSDs in the workplace, we present a video-based lifting action recognition method using rank-altered kinematic feature pairs, called top-scoring pairs (TSPs). We derive TSPs from a video dataset containing lifting and other activities commonly seen in MMH. These TSPs collectively classify each frame as lifting and non-lifting. The validation process involves evaluating classification performance. The proposed method minimizes computational and memory requirements while achieving performance comparable to more complex methods with greater computational demands. This makes it suitable for systems with limited hardware resources, thereby providing extensive feasibility across a variety of MMH environments to improve workplace safety.

RUSNet: Robust fish segmentation in underwater videos based on adaptive selection of optical flow

Fish segmentation in underwater videos can be used to accurately determine the silhouette size of fish objects, which provides key information for fish population monitoring and fishery resources survey. Some researchers have utilized underwater optical flow to improve the fish segmentation accuracy of underwater videos. However, the underwater optical flow is not evaluated and screen in existing works, and its predictions are easily disturbed by motion of non-fish. Therefore, in this paper, by analyzing underwater optical flow data, we propose a robust underwater segmentation network, RUSNet, with adaptive screening and fusion of input information. First, to enhance the robustness of the segmentation model to low-quality optical flow inputs, a global optical flow quality evaluation module is proposed for evaluating and aligning the underwater optical flow. Second, a decoder is designed by roughly localizing the fish object and then applying the proposed multidimension attention (MDA) module to iteratively recover the rough localization map from the spatial and edge dimensions of the fish. Finally, a multioutput selective fusion method is proposed in the testing stage, in which the mean absolute error (MAE) of the prediction using a single input is compared with that obtained using multisource input. Then, the information with the highest confidence is selected for predictive fusion, which facilitates the acquisition of the ultimate underwater fish segmentation results. To verify the effectiveness of the proposed model, we trained and evaluated it using a publicly available joint underwater video dataset and a separate DeepFish public dataset. Compared with the advanced underwater fish segmentation model, the proposed model has greater robustness to low-quality background optical flow in the DeepFish dataset, with the mean pixel accuracy (mPA) and mean intersection over union (mIoU) values reaching 98.77% and 97.65%, respectively. On the joint dataset, the mPA and mIoU of the proposed model are 92.61% and 90.12%, respectively, which are 0.72% and 1.21% higher than those of the advanced underwater video object segmentation model MSGNet. The results indicate that the proposed model can adaptively select the input and accurately segment fish in complex underwater scenes, which provides an effective solution for investigating fishery resources.

Remote Kinematic Analysis for Mobility Scooter Riders Leveraging Edge AI

Current kinematic analysis for patients with upper or lower extremity challenges is usually performed indoors at the clin- ics, which may not always be accessible for all patients. On the other hand, mobility scooter is a popular assistive tool used by people with mobility disabilities. In this study, we introduce a remote kinematic analysis system for mobility scooter riders to use in their local communities. In order to train the human pose estimation model for the kinematic anal- ysis application, we have collected our own mobility scooter riding video dataset which captures riders’ upper-body move- ments. The ground truth data is labeled by the collaborating clinicians. The evaluation results show high system accuracy both in the keypoints prediction and in the downstream kine- matic analysis, compared with the general-purpose pose mod- els. Our efficiency test results on NVIDIA Jetson Orin Nano also validate the feasibility of running the system in real-time on edge devices.

Disentangling Stochastic PDE Dynamics for Unsupervised Video Prediction.

Unsupervised video prediction aims to predict future outcomes based on the observed video frames, thus removing the need for supervisory annotations. This research task has been argued as a key component of intelligent decision-making systems, as it presents the potential capacities of modeling the underlying patterns of videos. Essentially, the challenge of video prediction is to effectively model the complex spatiotemporal and often uncertain dynamics of high-dimensional video data. In this context, an appealing way of modeling spatiotemporal dynamics is to explore prior physical knowledge, such as partial differential equations (PDEs). In this article, considering real-world video data as a partly observed stochastic environment, we introduce a new stochastic PDE predictor (SPDE-predictor), which models the spatiotemporal dynamics by approximating a generalized form of PDEs while dealing with the stochasticity. A second contribution is that we disentangle the high-dimensional video prediction into low-level dimensional factors of variations: time-varying stochastic PDE dynamics and time-invariant content factors. Extensive experiments on four various video datasets show that SPDE video prediction model (SPDE-VP) outperforms both deterministic and stochastic state-of-the-art methods. Ablation studies highlight our superiority driven by both PDE dynamics modeling and disentangled representation learning and their relevance in long-term video prediction.

Low-rank quaternion tensor completion for color video inpainting via a novel factorization strategy

Recently, a quaternion tensor product named Qt-product was proposed, and then the singular value decomposition and the rank of a third-order quaternion tensor were given. From a more applicable perspective, we extend the Qt-product and propose a novel multiplication principle for third-order quaternion tensor named gQt-product. With the gQt-product, we introduce a brand-new singular value decomposition for third-order quaternion tensors named gQt-SVD and then define gQt-rank and multi-gQt-rank. We prove that the optimal low-rank approximation of a third-order quaternion tensor exists and some numerical experiments demonstrate the low-rankness of color videos. So, we apply the low-rank quaternion tensor completion to color video inpainting problems and present alternating least-square algorithms to solve the proposed low gQt-rank and multi-gQt-rank quaternion tensor completion models. The convergence analyses of the proposed algorithms are established and some numerical experiments on various color video datasets show the high recovery accuracy and computational efficiency of our methods.

MWVOS: Mask-Free Weakly Supervised Video Object Segmentation via promptable foundation model

The current state-of-the-art techniques for video object segmentation necessitate extensive training on video datasets with mask annotations, thereby constraining their ability to transfer zero-shot learning to new image distributions and tasks. However, recent advancements in foundation models, particularly in the domain of image segmentation, have showcased robust generalization capabilities, introducing a novel prompt-driven paradigm for a variety of downstream segmentation challenges on new data distributions. This study delves into the potential of vision foundation models using diverse prompt strategies and proposes a mask-free approach for unsupervised video object segmentation. To further improve the efficacy of prompt learning in diverse and complex video scenes, we introduce a spatial–temporal decoupled deformable attention mechanism to establish an effective correlation between intra- and inter-frame features. Extensive experiments conducted on the DAVIS2017-unsupervised and YoutubeVIS19&21 and OIVS datasets demonstrate the superior performance of the proposed approach without mask supervision when compared to existing mask-supervised methods, as well as its capacity to generalize to weakly-annotated video datasets.

AVENUE: A novel deep fake detection method based on temporal convolution network & rPPG information

In Deep Learning (DL), an adversary creates Deepfakes by manipulating facial features to fool someone. The Deepfakes pose a security threat to anyone's privacy and a primary concern for our society. It can be detected by utilizing the texture and physiological properties of the face, like eye and lip movements; however, such methods are incompetent when Deepfakes are created using recent generative adversarial networks (GAN). Alternatively, remote Photoplethysmography (rPPG) information can be used for Deepfake detection because GANs neglect human physiological information for Deepfake generation. Such detection can be inaccurate when rPPG signals are affected by the noises induced by facial deformation and illumination variations. Furthermore, the exiting Deepfake detections are usually performed using sequential models, and such models fail to process the long sequence of temporal information. These issues are mitigated by our proposed method AVENUE , that is, \(A\) no \(V\) el d \(E\) epfake detectio \(N\) method based on temporal convol \(U\) tion n \(E\) twork & rPPG information. For mitigating the noise issues in the rPPG signals, the proposed method detects and employs relatively stable clips of the input video for Deepfake detection. The stable clips are those clips that are least affected by facial deformations. Also, we use a modified Temporal convolution network to model the long sequence of Deepfake information rather than the sequential architectures. We performed the experimental result on publically available datasets of Deepfake videos. It demonstrates that our proposed method performs better than the existing rPPG-based Deepfake detection methods.

Detecting clinical medication errors with AI enabled wearable cameras

Drug-related errors are a leading cause of preventable patient harm in the clinical setting. We present the first wearable camera system to automatically detect potential errors, prior to medication delivery. We demonstrate that using deep learning algorithms, our system can detect and classify drug labels on syringes and vials in drug preparation events recorded in real-world operating rooms. We created a first-of-its-kind large-scale video dataset from head-mounted cameras comprising 4K footage across 13 anesthesiology providers, 2 hospitals and 17 operating rooms over 55 days. The system was evaluated on 418 drug draw events in routine patient care and a controlled environment and achieved 99.6% sensitivity and 98.8% specificity at detecting vial swap errors. These results suggest that our wearable camera system has the potential to provide a secondary check when a medication is selected for a patient, and a chance to intervene before a potential medical error.

Reputation Burning: Analyzing the Impact of Brand Sponsorship on Social Influencers

The growth of the influencer marketing industry warrants an empirical examination of the effect of posting sponsored videos on influencers’ reputations. We collected a novel data set of user-generated YouTube videos created by prominent English-speaking influencers in the beauty and style category. We extracted a rich set of theory-driven video features and used DiNardo-Fortin-Lemieux reweighting to construct comparable treatment and control groups matched at the influencer-video level. The empirical analysis of the matched sample revealed a reputation-burning effect; that is, posting a sponsored video, compared with posting an equivalent organic video, cost the influencers 0.19% of their reputation (operationalized as the number of subscribers). The reputation-burning effect was stronger among the influencers with larger audiences. An analysis of likes, comments, and comment texts revealed a larger gap in audience response between sponsored and organic videos among the influencers with larger (versus smaller) audiences. The reputation-burning effect was mitigated when a high fit existed between the sponsored content and the influencer’s “usual” content and when the promoted brand was less well known. Our study empirically tested the assumption of several theoretical works. Moreover, it contributes to the literature on influencer marketing and celebrity endorsements and provides managerial implications for influencers, brands, and social media platforms. This paper was accepted by D. J. Wu, information systems. Funding: This research was supported by funding provided by Harvard Business School. Supplemental Material: The online appendix and data files are available at https://doi.org/10.1287/mnsc.2023.00193 .

Deep learning-based surgical step recognition for laparoscopic right-sided colectomy.

Understanding the complex anatomy and surgical steps involved in laparoscopic right-sided colectomy (LAP-RC) is essential for standardizing the surgical procedure. Deep-learning (DL)-based computer vision can achieve this. This study aimed to develop a step recognition model for LAP-RC using a dataset of surgical videos with annotated step information and evaluate its recognition performance. This single-center retrospective study utilized a video dataset of laparoscopic ileocecal resection (LAP-ICR) and laparoscopic right-sided hemicolectomy (LAP-RHC) for right-sided colon cancer performed between January 2018 and March 2022. The videos were split into still images, which were divided into training, validation, and test sets using 66%, 17%, and 17% of the data, respectively. Videos were manually classified into eight main steps: 1) medial mobilization, 2) central vascular ligation, 3) dissection of the superior mesenteric vein, 4) retroperitoneal mobilization, 5) lateral mobilization, 6) cranial mobilization, 7) mesocolon resection, and 8) intracorporeal anastomosis. In a simpler version, consecutive surgical steps were combined, resulting in five steps. Precision, recall, F1 scores, and overall accuracy were assessed to evaluate the model's performance in the surgical step classification task. Seventy-eight patients were included; LAP-ICR and LAP-RHC were performed in 35 (44%) and 44 (56%) patients, respectively. The overall accuracy was 72.1% and 82.9% for the eight-step and combined five-step classification tasks, respectively. The automatic surgical step-recognition model for LAP-RCs, developed using a DL algorithm, exhibited a fairly high classification performance. A model that understands the complex steps of LAP-RC will aid the standardization of the surgical procedure.

SiT-YOLOv9: An Efficient Algorithm for Learning Behavior Detection in the Home Environment

In the context of home-based learning, accurate identification of learning behaviors is essential for enhancing post-classroom learning efficiency. However, due to background interference and computational constraints in the TinyML terminal within home environments, CNN-based algorithms are susceptible to reduced performance and accuracy, leading to an increased false positive rate. To address this challenge, we propose a lightweight detection model called SiT-YOLOv9, which integrates MODNet, image enhancement, and other modules into the YOLOv9 model while also implementing moderate network pruning to effectively mitigate issues related to image noise and training sample computational power. Evaluation of the SiTBehaviors video dataset demonstrates that the SiT YOLOv9 model achieves outstanding performance with a recognition accuracy of 0.948 (mAP50) at a high processing speed of 90.9 frames per second. When compared with original models such as YOLOv8, YOLOv10, and RT-DERT, our proposed model exhibits superior recognition accuracy of 0.948 mAP, and a processing speed of 0.2 ms.

Welding Defect Monitoring Based on Multi-Scale Feature Fusion of Molten Pool Videos.

Real-time quality monitoring through molten pool images is a critical focus in researching high-quality, intelligent automated welding. However, challenges such as the dynamic nature of the molten pool, changes in camera perspective, and variations in pool shape make defect detection using single-frame images difficult. We propose a multi-scale fusion method for defect monitoring based on molten pool videos to address these issues. This method analyzes the temporal changes in light spots on the molten pool surface, transferring features between frames to capture dynamic behavior. Our approach employs multi-scale feature fusion using row and column convolutions along with a gated fusion module to accommodate variations in pool size and position, enabling the detection of light spot changes of different sizes and directions from coarse to fine. Additionally, incorporating mixed attention with row and column features enables the model to capture the characteristics of the molten pool more efficiently. Our method achieves an accuracy of 97.416% on a molten pool video dataset, with a processing time of 16 ms per sample. Experimental results on the UCF101-24 and JHMDB datasets also demonstrate the method's generalization capability.

A Novel DC GCN with Attention Mechanism for Accurate Near Duplicate Video Data Cleaning

There has been a steady emergence of nearly identical recordings in the last several decades, thanks to the exponential development of video data. The use of regular videos has been impacted by data quality difficulties produced by near-duplicate movies, which are becoming increasingly noticeable. While there has been progress in the field of near-duplicate video detection, there is still no automated merging method for video data characterised by high-dimensional features. As a result, it is challenging to automatically clean near-duplicate videos in advance video dataset data quality. Research on removing near-duplicate video data is still in its early stages. The precision of near-duplicate video data cleaning is severely compromised by the delicate issues of video data organization besides initial clustering centres in the current research, which arise when the previous distribution is unknown. In tackle these problems, we offer a new kind of Graph Convolutional Neural Network (GCN) that uses dense influences and a categorization attention mechanism. Deeply connected graph convolutional networks (DC-GCNs) learn about faraway nodes by making GCNs deeper. By using dense connections, the DC-GCN is able to multiplex the small-scale features of shallow layers and generate features at diverse scales. Finally, an attention mechanism is incorporated to aid in feature combination and importance determination. Sparrow Search Optimisation Algorithm (SSA) is used to pick the parameters of the given model in the most optimal way. In the end, experiments are carried out using a coal mining video dataset and a widely known dataset called CC_WEB_VIDEO. The simulation findings show that the suggested strategy performs better than certain previous studies.

Early fire detection using wavelet based features

In recent years, millions of hectares of vegetation worldwide have been destroyed by wildfires and forest fires. Computer vision-based fire classification, which separates pixels from image or video datasets into fire and non-fire categories, has gained more attention recently due to technological innovations. Fire pixels in an image or video can be classified using either a deep learning strategy or a traditional machine learning approach. Deep learning algorithms could process enormous volumes of data, but their training model performance is constrained because they fail to consider the differences in complexity between training samples. Moreover, it is not obvious to train a deep learning model without a dedicated GPU unit.Similarly, deep learning techniques that have a scarcity of training data and insufficient features exhibit poor performance in intricate real-world fire situations. Consequently, to categorize fire and non-fire pixels from the processed photos from the publicly accessible datasets, Corsican and FLAME, as well as the aerial private dataset Firefront_Gestosa, the current study uses a lightweight technique based on SVM and a refined set of features.The present research implemented a novel framework for fire detection and classification from a variety of RGB and Infra-red images acquired during real missions, addressing the significant requirement for swift and accurate recognition of diverse types of flames, ranging from wildfires to industrial and domestic fires. The framework employs wavelet decomposition-based features, including wavelet length, standard deviation, variance, energy, and Shannon’s entropy, extracted through a sliding window sampling method within a machine learning approach. It should be noted that managing multidimensional data to train a model is difficult in machine learning applications. This issue is solved adopting a feature selection approach, which eliminates redundant or unnecessary data that affects the functionality of the model. Thus, to enhance model performance, feature selection using ranking algorithms based on theoretical mutual information is applied in combination to the Support Vector Machine (SVM) with Radial Basis Function (RBF) kernel.Extensive experiments demonstrate exceptional results, notably with Haar wavelets, achieving an impressive overall accuracy of 99.43% and remarkable performance across specificity, precision, recall, F-measure, and G-mean metrics. Thus, the present study showcases the potential of advanced image processing techniques to significantly advance fire detection and classification, thereby contributing to fire prevention, management, and research in various contexts.