Abstract:

In recent years, there has been emphasis on museum learning because it provides visitors with the opportunity to experience culture and develop different interests. To improve the learning effect, it is necessary to enhance user experience. To do so, humans have traditionally guided visitors through the museum, but human resources are limited. Therefore, in recent years, augmented reality applications and robots that present information using deep learning have been attracting attention. These use deep learning to detect exhibits and provide information about them to users. This information enables users to deepen their understanding of the exhibits, and thus improves user experience. However, in a museum, the lighting is suppressed to preserve the exhibits, and it is difficult to detect exhibits in the dark because the captured images become unclear. Therefore, it is necessary to enhance the low-light images for object detection in the dark.
In this paper, we propose a low-light image enhancement technique using image-to-image translation that can be trained without pair data. Image-to-image transformation is a technique that transforms a set of images with one common feature into a set with another feature. Our method enhanced the low-light images by training to transform them from "low-light images" to "normal-light images". Unsupervised learning is possible by retransforming the image during training to preserve the shape. To prevent color changes due to the transformation, we employed a loss function for color preservation before and after the transformation. To evaluate the proposed method, we investigated the performance of the object detection task under low-light conditions. First, we photographed low-light images under five different lighting conditions. Next, we performed a low-light image enhancement task for the photographed images. Then, we evaluated the detection rate by detecting the objects in these images. The experiments demonstrate that the proposed method outperforms the conventional method in terms of the detection rate.

Keywords:

Low-light image enhancement, Object detection.