Objective Night-vision infrared imaging technology plays a crucial role in fields such as intelligent driving. However, in nighttime low-temperature environments, infrared images influenced by the urban heat island effect and seasonal cold is characterized by overall low brightness, darker backgrounds, and severe degradation of target details. This degradation impacts the precise perception for road condition by intelligent driving system. Consequently, it is urgent to establish an effective infrared image enhancement method for low-temperature scenes, to improve the overall image brightness and detail presentation. Traditional infrared image enhancement methods rely on fixed rules and lack in depth understanding of the image semantic information, resulting in poor generalization. At present, deep learning-based methods have been widely adopted. However, most of these approaches rely on supervised learning, which is constrained by the difficulty of obtaining paired high-quality object images for low-temperature infrared images. Unsupervised learning-based methods eliminate the dependency on paired images by learning the distribution characteristics of image data, which demonstrate excellent generalization. Nevertheless, the existing approaches tend to produce overexposure or underexposure when dealing with nighttime low-temperature infrared image. Therefore, a Deep Network-Estimated Adaptive Curve Enhancement Network (DNE-ACENet) has been proposed.

Methods The DNE-ACENet employs iterative optimization through a designed nonlinear cubic brightness enhancement curve, progressively boosting brightness and detail in dark areas while preventing local overexposure in highlights. Parameters of the curve cluster are generated by the Parameter Estimation Block (PEB) according to the mapping relationship between the key features of original image and the curve parameters. The core of the PEB is that utilizes a Transformer-Wavelet hybrid structure to establish long-range dependencies at different scales, enabling multi-scale information fusion and reconstruction for global low-frequency feature enhancement. Additionally, a densely connected mechanism with small receptive fields is adopted to focus on detailed information, preserving local features through feature reuse, thereby achieving precise capture and expression of local high-frequency details. Furthermore, the contrast loss function is incorporated into the multi-loss collaborative optimization strategy, to enhance image contrast in low-temperature and increase the distinguishability between the target and the background.

Results and Discussions The comparative experiments demonstrate that the visual effect of DNE-ACENet is superior to that of the five existing state-of-the-art networks on public datasets and images collected from actual roads. The values of CTR, IE, CEIQ and NIQE of DNE-ACENet are all optimal, and improved by at least 5.26%, 2.28% and 2.63% respectively. The value of μ is suitable for visual observation (Tab.1). The overall brightness of the enhanced image is significantly improved, and the target is easier to identify (Fig.1), indicating that DNE-ACENet is more suitable for infrared image enhancement in low-temperature scenes at night. The ablation experiment shows that the synergistic effect of nonlinear cubic curves, Transformer-Wavelet hybrid structures, and dense connection mechanisms can achieve a reasonable improvement in brightness while ensuring contrast, detail richness, and image quality (Tab.2). The network efficiency evaluation indicates that the computational complexity of DNE-ACENet is 5.67G FLOPs, and the parameter number is 0.11 M, which is comparable to the lightweight methods. Moreover, its Frame rate is 158.73 FPS, which can meet the real-time requirements (Tab.3). Furthermore, the target detection experiments show that the images enhanced by DNE-ACENet increase the AP_Peds, AP_Car and mAP of the detection model by at least 3.18%, 1.71% and 3.92% respectively, enabling more accurate identification (Tab.4). These results validate that DNE-ACENet is more suitable for enhancing infrared images in nighttime low-temperature scenes.

Conclusions The DNE-ACENet is applicable to enhancing nighttime low-temperature infrared images. The Iterative Enhancement Block (IEB) achieves a progressive improvement for the brightness and details in the dark areas while suppresses local overexposure in the bright areas through a cluster of nonlinear cubic curves. The PEB enhances the key image features through the Transformer-Wavelet hybrid structure and the dense connection mechanism, and strengthens the mapping relationship between image features and curve parameters to generate the optimal parameters for the curves. The experimental results demonstrate that the DNE-ACENet is superior to the existing methods in improving the brightness, detail, contrast, and the overall visual effect of nighttime low-temperature infrared images. This study provides an effective technical solution for environmental perception in intelligent driving systems under nighttime low-temperature conditions, offering reliable support for downstream tasks such as target detection and tracking, as well as path planning.