Open AccessSkip-layer network with optimization method for domain adaptive detection
Author(s) -
Qian Xu,
Ying Li,
Gang Wang,
Mengshu Hou,
Hao Zhang,
Hongmin Cai
Publication year - 2022
Publication title -
plos one
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.99
H-Index - 332
ISSN - 1932-6203
DOI - 10.1371/journal.pone.0263748
Subject(s) - computer science , artificial intelligence , classifier (uml) , domain (mathematical analysis) , pattern recognition (psychology) , convolutional neural network , object detection , effective domain , computer vision , algorithm , convex optimization , mathematics , convex combination , regular polygon , mathematical analysis , geometry
In the field of object detection, domain adaptation is one of popular solution to align the distribution between the real scene (target domain) and the training scene (source domain) by adversarial training. However, only global features are applied to the Domain Adaptive Faster R-CNN (DA Faster R-CNN) method. The lack of local features reduces the performance of domain adaptation. Therefore, a novel method for domain adaptive detection called Skip-Layer Network with Optimization (SLNO) method is proposed in this paper. Three improvements are presented in SLNO. Firstly, different level convolutional features are fused by a multi-level features fusion component for domain classifier. Secondly, a multi-layer domain adaptation component is developed to align the image-level and the instance-level distributions simultaneously. Among this component, domain classifiers are used in both image-level and instance-level distributions through the skip layer. Thirdly, the cuckoo search (CS) optimization method is applied to search for the best coefficient of SLNO. As a result, the capability of domain alignment is strengthened. The Cityscapes, Foggy Cityscapes, SIM10K, KITTI data sets are applied to test our proposed novel approach. Consequently, excellent results are achieved by our proposed methods against state-of-the-art object detection methods. The results demonstrate our improvements are effective on domain adaptation detection.