Enhancing Dynamic Images For End-To-End Training-Based Object Detection
Dharmesh Dhabliya, K. S. Bhuvaneshwari, Ashmeet Kaur, G. Ramya, Sourav Rampal, S.E Manu · 2024
In computer vision, object identification using convolutional neural networks is currently trending. There will be a dramatic drop in detection ability in low-light situations due to the image's lighting component, which severely affects object identification. For improved detection results, it is recommended to use a low light picture enhancement approach as a pre processing mechanism to increase image quality. However, current enhancing strategies could be ineffective on some samples because of the intricacy of low-light conditions. As a result, doing better overall detection in low-light circumstances is challenging. We focus on improving object recognition accuracy using picture augmentation in this work, not human perceptual quality. The proposed image enhancement model is compared with the other enhancement techniques such as Histogram equalization (HE), Contrast limited adaptive histogram equalization (CLAHE), gamma correction (GC), and balance contrast enhancement technique (BCET). For the object detection we have used three different methods of CNN to evaluate our image enhancementmodel. With a high detection rate of 98%, the suggested technique outperforms competing CNN algorithms on bigger datasets. When compared to GC, HE, CLAHE, and BCET, our suggested strategy yields more efficient object identification with values of 0.8219), 0.8052 (recall), and 0.8133 (F1-score).