Integration of Single Shot Multibox Detector and Single Scale Retinex in Detecting Oil Palm Fruit Bunches in Low Light Conditions

Abstract

This research develops a computer vision-based system to detect oil palm fruit bunches in low light conditions. The main contribution of this study is to integrate two methods that can be used to detect oil palm bunches in low light conditions caused by cloudy weather conditions and dense foliage. The proposed approach is to integrate Single Scale Retinex (SSR) to improve the quality of light intensity in the image, then the detection process is carried out by Single Shot Multibox Detector (SSD). The experiment was carried out using images of oil palm fruit bunches taken directly on the tree using a resolution of 1920x1080 pixels and a shutter speed of 1/400 and ISO 400. The tests were divided into three categories based on light intensity level (lux) and data acquisition time, each consisting of 104 images for 245 bunches. In the first category (1000–1600 lux, daytime), the detection results achieved 100% precision, 85.31% recall, and 92.07% F1 score. The second category (500–999 lux, morning) recorded a precision score of 99.52%, a recall of 84.08%, and an F1 score of 91.15%. Meanwhile, the third category (200–499 lux, afternoon) produced a precision of 98.55%, a recall of 84.30%, and an F1 score of 90.87%. A comparative analysis with ISO 3200 camera settings without an image enhancement algorithm shows that the combination of ISO 400 and SSR provides a 16.75% performance improvement at 1000–1600 lux light intensity, by 48.2% at 500–999 lux light intensity and 69.21% at 200–499 lux and. A comparison was also made with SSDs integrated with MSR, the results showed that SSDs integrated with SSR had a high F1 Score. These results confirm that integrating SSR with the detection model can improve the detection performance of the model in low and uneven light conditions.

Access to Document

Other files and links

• Link to publication in Scopus

Fingerprint