EVALUASI ROBUSTNESS DETEKSI OBJEK PADA CITRA TERDEGRADASI ASAP MENGGUNAKAN YOLO DENGAN PIPELINE MULTI-TAHAP CLAHE–WIENER
Description
Object detection in images that have undergone degradation, especially due to smoke interference, has become a crucial issue in the field of computer vision. This condition generally causes a decrease in contrast as well as an increase in noise, which significantly impacts the performance of detection models. To address this problem, this study proposes a multi-stage image preprocessing approach that integrates the Contrast Limited Adaptive Histogram Equalization (CLAHE) method with the Wiener filter. This approach is aimed at improving the robustness of YOLO-based object detection models in facing extreme degradation conditions. The main focus of the research contribution lies in the design of a CLAHE-Wiener tiered preprocessing pipeline applied before YOLO inference, as well as the development of a quantitative evaluation scheme using the Mean Average Precision (mAP) metric at various levels of smoke interference. It should be emphasized that this study does not claim novelty in the basic algorithm, considering that CLAHE and the Wiener filter have been widely known in previous literature. The experimental process was carried out by generating data variations through the addition of synthetic noise to images within the intensity range of 0% to 90%. Subsequently, the degraded images were processed using a combination of CLAHE and the Wiener filter gradually before being input into the YOLO model for object detection. Evaluation was conducted using mAP, precision, and recall metrics on a total of 1,020 images, divided into 900 training data, 300 validation, and 300 testing. The test results indicated that the application of this multi-stage pipeline provided significant performance improvement, especially at noise levels above 60%. Under the highest noise condition (90%), the YOLO model without preprocessing only achieved an mAP of 61.0, while the proposed approach was able to increase this value up to 77.0. In addition, the stability of precision and recall values also tends to be better maintained under high degradation conditions after the preprocessing stage is applied. Overall, the combination of CLAHE and the Wiener filter in a multi-stage scheme has been proven effective in improving image quality while strengthening the resilience of object detection systems against severe visual disturbances such as smoke. Nevertheless, the results obtained are still limited to scenarios with synthetic noise, so further testing using real-world data is needed to validate the reliability of the method more comprehensively
Files
Institutions
- State University of MalangEast Java, Malang