An Attention Enhanced Bidirectional LSTM for Early Forest Fire Smoke Recognition

Abstract

Detecting forest fire smoke during the initial stages is vital for preventing forest fire events. Recent studies have shown that exploring spatial and temporal features of the image sequence is important for this task. Nevertheless, since the long distance wildfire smoke usually move slowly and lacks salient features, accurate smoke detection is still a challenging task. In this paper, we propose a novel Attention Enhanced Bidirectional Long Short-Term Memory Network (ABi-LSTM) for video based forest fire smoke recognition. The proposed ABi-LSTM consists of the spatial features extraction network, the Bidirectional Long Short-Term Memory Network(LSTM), and the temporal attention subnetwork, which can not only capture discriminative spatiotemporal features from image patch sequences but also pay different levels of attention to different patches. Experiments show that out ABi-LSTM is capable of achieving best accuracy and less false alarms on different types of scenarios. The ABi-LSTM model achieve a highly accuracy of 97.8%, and there is 4.4% improvement over the image-based deep learning model.

Two-Step Real-Time Night-Time Fire Detection in an Urban Environment Using Static ELASTIC-YOLOv3 and Temporal Fire-Tube

“…Precision (%) Recall (%) F1-Score (%) In addition, we compared the fire candidate detection performance and processing time per frame with the Faster R-CNN- [33] and single-shot detector (SSD)-based fire detectors, which are widely used in object detection and other fire detection studies. Table 3 shows the comparison results of the four methods.…”

“…However, because fire detection is applied based on a CNN with a still image without considering the temporal variation of the flame, such methods have a high probability of a false detection of the surrounding fire-like objects as an actual fire. To solve this problem, some methods [32][33][34] have combined an RNN or an LSTM with a CNN when considering the spatio-temporal characteristics of the sequential fire flames. With these approaches, a CNN is normally used to extract the spatial features, and an RNN or LSTM is used to learn the temporal relation between frames.…”

“…Han et al [32] combined a CNN with an RNN in a consecutive manner such that the sequence data could be allowed in the model. Cao et al [33] proposed an attention-enhanced bidirectional LSTM for camera-based forest fire recognition. Kim and Lee [34] used Faster R-CNN to detect the suspected regions of a fire and then summarized the features within the bounding boxes in successive frames accumulated by the LSTM to classify whether a fire was present within a short-term period.…”

“…Because action recognition requires a spatiotemporal feature analysis similar to fire, 3D ConvNets are combined with ELASTIC-YOLOv3 (ELASTIC-YOLOv3 + 3D ConvNets [48]). In the case of LSTM, we adopted the method in [33] to apply a feature vector converted from a CNN to the LSTM layer in a frame-by-frame manner and verify the fire through a continuous LSTM output. Like other comparison methods, only frames detected by ELSATIC-YOLOv3 as fire candidates were input into the LSTM (ELASTIC-YOLOv3 + LSTM [33]).…”

“…In the case of LSTM, we adopted the method in [33] to apply a feature vector converted from a CNN to the LSTM layer in a frame-by-frame manner and verify the fire through a continuous LSTM output. Like other comparison methods, only frames detected by ELSATIC-YOLOv3 as fire candidates were input into the LSTM (ELASTIC-YOLOv3 + LSTM [33]). All experiments were measured after 50 frames because fire verification was required only after a certain frame had been accumulated.…”