TL;DR
This paper presents a high-efficiency edge detection model that combines an improved Sobel algorithm with YCbCr color space conversion, specifically optimized for FPGA hardware. By expanding the Sobel operator to a 5x5 kernel with eight detection directions and leveraging the parallel processing power of FPGAs, the researchers achieved superior accuracy (0.90) and significantly lower power consumption (4W) compared to traditional CPU and GPU implementations.
What Problem Does This Paper Address?
Edge detection is a fundamental building block in computer vision, essential for tasks like object recognition and medical image analysis. However, traditional algorithms often face a "trilemma": they struggle to balance real-time processing speed, computational efficiency, and accuracy in noisy environments.
Standard Sobel operators typically use a 3x3 kernel that only looks at horizontal and vertical gradients, which can miss fine details or produce "false edges" when noise is present. Furthermore, performing these calculations on general-purpose CPUs or power-hungry GPUs is often impractical for edge-computing scenarios (like autonomous driving or IoT devices) where low latency and low power consumption are non-negotiable.
Methods
The authors propose a multi-stage pipeline optimized for hardware acceleration:
- Keyframe Extraction: Using Optical Flow (OF) to identify frames with the highest information content from video streams, ensuring the model processes the most relevant data.
- YCbCr Color Space Conversion: Instead of working in standard RGB, the model converts images to YCbCr. Since the human eye (and edge detection math) is more sensitive to brightness (Luminance, or the 'Y' component) than color, the model focuses on the Y component to enhance edge clarity.
- Improved Sobel Operator: The core innovation is moving from a standard 3x3 template to a 5x5 convolution kernel with eight detection directions (0°, 22.5°, 45°, etc.). This allows the model to capture much more complex structural information.
- Adaptive Thresholding (Otsu Method): To distinguish edges from background noise, the model uses the Otsu method to dynamically calculate the optimal threshold for each image.
- FPGA Optimization: The entire pipeline was implemented on a Xilinx Zynq UltraScale+ MPSoC. The authors utilized Block RAM (BRAM) for high-speed data access and a pipeline architecture to ensure that each operation executes within a single clock cycle.
Datasets
The researchers evaluated their model using the Kaggle Surveillance Video Dataset.
- Composition: 6,000 images extracted from surveillance footage.
- Variety: Includes various resolutions (VGA, 720p, 1080p) and diverse subjects like pedestrians and vehicles.
- Conditions: The dataset covers both daytime and nighttime scenes to test the model's robustness under different lighting conditions.
Key Insights and Ideas
- Luminance over Color: By isolating the 'Y' component in the YCbCr space, the model reduces redundant data processing while actually improving the "sharpness" of the detected edges.
- Directional Sensitivity: Moving from two directions (horizontal/vertical) to eight directions significantly reduces the "jaggedness" of diagonal edges, which is a common failure point for simpler algorithms.
- Hardware-Algorithm Co-design: The choice of a 5x5 kernel would typically be too computationally expensive for real-time software. However, because FPGAs can perform these convolutions in parallel using dedicated hardware logic, the increased complexity results in better accuracy without a proportional hit to latency.
Key Results
The proposed FPGA-IpSobel model outperformed standard Roberts and Sobel operators across several metrics:
- Accuracy & Quality: At a sample size of 100, the model achieved an accuracy (ACC) of 0.90 and a low Root Mean Square Error (RMSE) of 0.16.
- Latency: The FPGA implementation clocked in at just 15ms per frame, compared to 28ms for a GPU (NVIDIA RTX 2080) and 45ms for a CPU (Intel i7-9700K).
- Power Efficiency: This is the standout result—the FPGA consumed only 4W, whereas the GPU and CPU consumed 75W and 65W, respectively. This represents a massive leap in performance-per-watt, making it ideal for embedded systems.
- Information Retention: The preprocessing model maintained an information retention rate of 0.89 with a structural loss of only 0.05.
Limitations and Future Work
Despite the performance gains, the authors note that the model still struggles in specific high-complexity scenarios:
- High-Frequency Textures: In environments like dense forests or grasslands, the adaptive thresholding may misidentify intricate textures as edges, leading to "false positives."
- Overlapping Objects: Distinguishing between multiple overlapping objects remains a challenge for this specific gradient-based approach.
Future Directions: The authors suggest developing a more robust fusion algorithm specifically for textured backgrounds and implementing a "scene-based threshold self-learning" system that can automatically adjust parameters based on the environment (e.g., switching modes between a city street and a forest).
Summary updated