Improving the Accuracy of Object Detection in Low Light Conditions using Multiple Retinex Theory-based Image Enhancement Algorithms

Object Detection is a vast field that has many applications in present and upcoming technologies. However, improving the accuracy of object detection algorithms remains a persistent challenge. There are some limitations to its accuracy and many factors like image quality, noise, and the illumination of the image play a crucial role in it. It is more likely that an image would have noise if it was captured in low illumination conditions as the camera captures less light. To solve this problem and improve object detection accuracy, this study proposes to pass the low exposure image through existing Retinex theory-based low light image enhancement models and then its output to be passed into an object detection algorithm. Retinex based image enhancement models estimate the areas with low exposures and noise is reduced from the image as well with the help of neural networks. This demonstrates a positive impact on the confidence values of the object detection and more tendency for an object to be detected. Lastly, a comparison has also been performed on three existing low light image enhancement models. MIRNet, MBLLEN, and TCN models have been used for comparison based on confidence values of the objects detected in various images.