Real-time infrared image edge enhancement and AI-enabled recognition using FPGA-based edge computing architecture

Abstract

Infrared images frequently exhibit low contrast, weak target boundaries, nonuniform detector responses, and scene-dependent thermal interference, making reliable edge extraction difficult when a fixed detector and thresholding rule are used. This study presents an AI-guided infrared edge-enhancement framework with a real-time field-programmable gate array (FPGA) streaming architecture. The method evaluates Sobel, Canny, Laplacian-of-Gaussian (LoG), and difference-of-Gaussian operators in combination with seven thresholding strategies. Misclassification error (ME) and edge mismatch measure (EMM) are combined into a composite score to assess pixel agreement and contour fidelity. These evaluation results provide the basis for a lightweight AI closed loop that analyzes image statistics and selects an appropriate thresholding policy according to the operating scene. Experiments were conducted using 1,341 infrared images acquired under sunny, cloudy, rainy, nighttime, and indoor conditions. Quantitative evaluation of representative labeled images shows that the preferred thresholding strategy is scene-dependent: Kapur achieves the best sunny-scene score of 0.8680, Otsu produces the best scores of 0.8873, 0.8887, and 0.8913 for cloudy, rainy, and nighttime scenes, respectively, and K-means achieves 0.8436 indoors. The edge-processing pipeline was implemented on a Cyclone V SoC FPGA connected to a 320 × 256 InSb focal-plane array. Gaussian, Sobel, LoG, zero-crossing, histogram, and fusion operations are executed using parallel fixed-point pipelines. The architecture supports a 5-MHz input pixel stream, while an eight-bit coarse-search Otsu module operating at 100 MHz reduces threshold-search time to approximately 0.655 ms. The proposed approach combines explainable edge processing, scene-dependent adaptability, and deterministic FPGA execution for embedded infrared imaging applications.