Lightweight multi-scale network for small object detection

• Published in: PeerJ. Computer science Vol. 8; p. e1145
• Main Authors: Li, Li, Li, Bingxue, Zhou, Hongjuan
• Format: Journal Article
• Language: English
• Published: San Diego PeerJ. Ltd 08-11-2022; PeerJ, Inc; PeerJ Inc
• Subjects: Accuracy; Algorithms; Algorithms and Analysis of Algorithms; Artificial Intelligence; Channel attention; Computer Vision; Datasets; Deep learning; Feature extraction; Lightweight; Machine learning; Methods; Modules; Multi-scale feature fusion; Multimedia; Neural Networks; Object recognition; Object recognition (Computers); Pattern recognition; Receptive field enhancement; Semantics; Small object detection; China
• ISSN: 2376-5992; 2376-5992
• DOI: 10.7717/peerj-cs.1145

Small object detection is widely used in the real world. Detecting small objects in complex scenes is extremely difficult as they appear with low resolution. At present, many studies have made significant progress in improving the detection accuracy of small objects. However, some of them cannot balance the detection speed and accuracy well. To solve the above problems, a lightweight multi-scale network (LMSN) was proposed to exploit the multi-scale information in this article. Firstly, it explicitly modeled semantic information interactions at every scale via a multi-scale feature fusion unit. Secondly, the feature extraction capability of the network was intensified by a lightweight receptive field enhancement module. Finally, an efficient channel attention module was employed to enhance the feature representation capability. To validate our proposed network, we implemented extensive experiments on two benchmark datasets. The mAP of LMSN achieved 75.76% and 89.32% on PASCAL VOC and RSOD datasets, respectively, which is 5.79% and 11.14% higher than MobileNetv2-SSD. Notably, its inference speed was up to 61 FPS and 64 FPS, respectively. The experimental results confirm the validity of LMSN for small object detection.