In recent years, deep neural networks have driven notable advancements in image recognition. However, achieving high recognition accuracy under low-light conditions remains a challenging issue. This paper presents a method aimed at improving recognition performance in such environments. We introduce an image-adaptive learnable module that applies tailored image processing to input images, along with a parameter predictor that estimates optimal image correction parameters for the module. Our method enhances recognition accuracy under low-light conditions by acting as a front-end filter that can be seamlessly integrated without the need to retrain existing models. We adopt two different pose estimation models as the recognition model and demonstrate through experiments that applying our method to these models leads to improved recognition accuracy.