Zhengqi Ding ¹ Kelei Sun ² Ting Li ³ Gang Sun ^1*

• ¹ Fuyang Normal University, Fuyang, China
• ² Anhui University of Science and Technology, Huainan, Anhui, China
• ³ Anhui Lanzhi Big Data Technology Co., Ltd., Hefei, China

Conventional object detection algorithms are limited by the availability of training and testing samples in the target domain. To alleviate the dependency on the target domain and reduce the burden of annotation while ensuring detection performance, this paper proposes the use of an IF-Net (Information Invariant Scaling and Feature Alignment Module Net) framework to extract feature information from unconstrained objects. Firstly, an Information Invariant Scaling (IIS) Module is proposed, which introduces multiple layers of information at different depths of the network, greatly improving the detection performance of the network. Secondly, a Feature Alignment Module (FAM) is introduced, which trains target boxes using samples from the source domain and aligns different features to enhance the network's performance and reduce its reliance on a large number of target domain samples. Finally, a novel loss function, ℒ IF , is proposed to further enhance the detection performance and accuracy of the network. Experimental studies conducted on three publicly available datasets demonstrate significant advantages of our framework compared to state-of-the-art algorithms.