Deep-Learning Smoke Detection With Color Transfer and Feature Calibration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional smoke detection methods face challenges in environments with high sensitivity requirements, complexity, and large detection ranges, particularly due to poor robustness, limited detection accuracy for small or distant smoke, and high false alarm rates due to self-similarity and transparency of smoke.
Innovation Solution
A smoke detection method based on deep learning that includes image acquisition, enhancement, color transfer, feature extraction, and multi-layered network processing to improve smoke data collection and detection accuracy, addressing issues of data acquisition and feature overlap.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional visual methods focus on learning color and texture features of smoke, then detection range and response speed are guaranteed, but robustness is poor and detection accuracy is limited for small or distant smoke
Solution Approach 1:
The patent segments the smoke detection task into multiple stages: initial detection using color/texture features for broad coverage, followed by progressive refinement through multiple detection rounds. Each round focuses on different feature aspects, with later rounds correcting errors from earlier rounds, thereby improving both accuracy and robustness without sacrificing detection range.
Solution Approach 2:
The patent introduces temporal dimension by performing multiple detection rounds on the same image sequence. Instead of relying on a single-frame analysis, the system accumulates detection results across multiple frames and rounds, adding a time-based dimension to the detection process that enhances robustness while maintaining real-time performance.
2Measurement precision
If deep learning is applied to smoke detection, then detection accuracy can be improved, but data acquisition is difficult due to safety hazards and smoke's self-similarity and transparency cause high false alarm rates
Solution Approach 1:
The patent performs preliminary data preparation by pre-processing images to enhance smoke features before feeding them to the deep learning model. This includes color space conversion, feature enhancement, and pre-detection filtering that reduces the complexity of raw data and improves model training efficiency, thereby reducing overall system complexity while maintaining high accuracy.
Solution Approach 2:
The patent introduces intermediate processing layers between raw image input and the deep learning model, including feature extraction modules and pre-processing steps that act as intermediaries. These intermediaries transform complex raw data into more manageable feature representations, reducing the burden on the main detection model and simplifying the overall data collection and processing pipeline.
3Area of stationary object
If smoke detection is performed through low-level features like color and texture, then large pieces of nearby smoke can be detected, but false alarm rate increases due to smoke's self-similarity and transparency
Solution Approach 1:
The patent implements periodic detection rounds where the same image is analyzed multiple times with different feature focus. Each detection round uses slightly different parameters and feature weights, and results are aggregated over time. This periodic re-detection reduces false alarms by requiring consistent detection across multiple rounds while maintaining the ability to detect large smoke areas.
Solution Approach 2:
The patent incorporates feedback mechanisms where detection results from one round inform the parameters and focus of subsequent rounds. High-confidence detections are reinforced, while ambiguous cases are re-examined with adjusted parameters. This feedback loop continuously refines detection accuracy and reduces false alarms while preserving detection range for large smoke areas.
Data Source
AI summary
The present application relates to a smoke detection method based on deep learning, a device and a storage medium. The method includes the following steps: performing image enhancement processing and color transfer processing on a smoke image to obtain a color transferred smoke image; superimposing the color transferred smoke image on an indoor image to obtain an initial image, and performing screening processing, detection frame updating processing, feature extraction and layer calibration processing on the initial image to obtain a smoke calibrated feature image; performing segmentation processing, detection frame prediction processing, detection frame classification processing and image classification processing on the smoke calibrated feature image in sequence to obtain a smoke image set; and screening the smoke image set to obtain a target smoke image, and outputting a smoke detection result. The present disclosure realizes the collection of smoke data sets and improves the detection accuracy of smoke.