Switching Neural Network For Improving The Accuracy Of Object Detection In Mixed Low-Light And Normal-Light Conditions

Mahendra, Rama Yusuf (2023) Switching Neural Network For Improving The Accuracy Of Object Detection In Mixed Low-Light And Normal-Light Conditions. Masters thesis, Institut Teknologi Sepuluh Nopember.

[thumbnail of 6022212008-Master_Thesis.pdf] Text
6022212008-Master_Thesis.pdf - Accepted Version
Restricted to Repository staff only until 1 October 2025.

Download (1MB) | Request a copy


Nowadays, object detection has been widely employed in various sectors. However, object detection performance at night is constrained by low-light conditions. Due to dark conditions, low-light images possess several degrading qualities, such as low visibility, low contrast, and large noise. The features also shift compared to the normal-light conditions. This condition causes the accuracy of object detection to decrease drastically. To overcome this problem, several studies have carried out methods, one of which is combining networks of low-light image enhancement and object detection into one. However, merging the two networks will result in each network cannot achieve its best function because it must accommodate other functions. In addition, the network will be large and difficult to train. Therefore, in this study, we propose to use a switching neural network to direct the image to an object detection model that has been trained more specifically based on light conditions. This switch is a two-class image classification model that will predict whether an image is low-light or normal-light. We use some variations on the switch model, such as VGG16 and EfficientFormerV2. The experimental results reveal that the combined switching neural network with YOLOv8 obtains superior overall mAP result compared to YOLOv8 without switching neural network which is only trained on a low-light or normal-light dataset and even surpassed YOLOv8 without switching neural network which is trained on a combined low-light and normal-light dataset at the score of 0.691, 0.632, 0.652, 0.569, respectively. For speed testing, it was found that the switching neural network + YOLOv8 made the processing time longer, approximately 2x for the VGG16 switch, 4.7x for the EfficientFormerV2 L switch, and 4.3x for the EfficientFormerV2 S2 switch.
Saat ini, deteksi objek telah banyak digunakan di berbagai sektor. Namun, performa pendeteksian objek di malam hari terkendala oleh kondisi cahaya rendah. Karena kondisi gelap, gambar dengan cahaya rendah memiliki beberapa kualitas yang menurun, seperti rendahnya visibilitas, rendahnya kontras, dan noise yang besar. Fiturnya juga bergeser dibandingkan dengan kondisi cahaya normal. Kondisi tersebut menyebabkan akurasi pendeteksian objek menurun drastis. Untuk mengatasi masalah tersebut, beberapa penelitian telah mengusulkan metode-metode, salah satunya adalah menggabungkan jaringan perbaikan gambar cahaya rendah dan pendeteksi objek menjadi satu. Namun, menggabungkan kedua jaringan tersebut akan menmbuat masing-masing jaringan tidak dapat mencapai fungsi terbaiknya karena harus mengakomodasi fungsi lainnya. Selain itu, jaringannya akan besar dan sulit untuk dilatih. Oleh karena itu, dalam penelitian ini kami mengusulkan untuk menggunakan switching neural network untuk mengarahkan gambar ke model pendeteksian objek yang telah dilatih lebih spesifik berdasarkan kondisi cahaya. Switching ini adalah model klasifikasi gambar dua kelas yang akan memprediksi apakah suatu gambar memiliki cahaya rendah atau cahaya normal. Kami menggunakan beberapa variasi model switching, seperti VGG16 dan EfficientFormerV2. Hasil percobaan menunjukkan bahwa kombinasi switching neural network dengan YOLOv8 memperoleh hasil mAP keseluruhan yang lebih unggul dibandingkan dengan YOLOv8 tanpa switching neural network yang hanya dilatih pada dataset cahaya rendah atau cahaya normal dan bahkan melampaui YOLOv8 tanpa switching neural network yang dilatih pada gabungan data cahaya rendah dan cahaya normal dengan skor 0,691, 0,632, 0,652, 0,569 secara berurutan. Untuk pengujian kecepatan, ditemukan bahwa switching neural network + YOLOv8 membuat waktu pemrosesan lebih lama, kira-kira 2x untuk switch VGG16, 4,7x untuk switch EfficientFormerV2 L, dan 4,3x untuk switch EfficientFormerV2 S2.

Item Type: Thesis (Masters)
Uncontrolled Keywords: image classification, low-light images, normal-light images, object detection, switching neural network, deteksi objek, gambar cahaya normal, gambar cahaya rendah, klasifikasi gambar
Subjects: T Technology > TA Engineering (General). Civil engineering (General) > TA1637 Image processing--Digital techniques
T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK5103.8 Switching systems
Divisions: Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Electrical Engineering > 20101-(S2) Master Thesis
Depositing User: Rama Yusuf Mahendra
Date Deposited: 20 Jul 2023 04:50
Last Modified: 20 Jul 2023 04:50
URI: http://repository.its.ac.id/id/eprint/98677

Actions (login required)

View Item View Item