MobileDenseNet: A new approach to object detection on mobile devices. (1st April 2023)

• Record Type: Journal Article
• Authors: Hajizadeh, Mohammad; Sabokrou, Mohammad; Rahmani, Adel
• Is Part Of: Expert systems with applications. Volume 215(2023)
• Publication Date: 2023-04-01
• Subjects: Object detection -- Embedded system -- Mobile device -- Deep neural network; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2022.119348
• Languages: English
• ISSNs: 0957-4174

Abstract

Abstract: Object detection problem-solving has witnessed vast developments in the past few years. There is a need for lighter models in cases where hardware limitations exist, together with a demand for models to be designed to run on mobile devices. This paper assesses methods used for algorithms which are created to address these issues. The main goal of this article is to improve accuracy in state-of-the-art algorithms while maintaining speed and real-time efficiency. Small objects and inaccurate localization are the main issues in one-stage object detection. To resolve these issues, in this paper, a new network entitled "MobileDenseNet" was created, which is suitable for embedded systems. We also developed a light neck, called "FCPNLite" for mobile devices, which facilitates the detection of small objects. Our research suggests that necks in embedded systems have been sporadically mentioned by the existing studies. The novelty of our network lies in the use of concatenation features. A small yet significant change to the head of the network improved accuracy without increasing speed or limiting parameters. Overall, the accuracy of our method on the challenging CoCo and Pascal VOC datasets were 24.8 and 76.8 percentage, respectively. This has been by far the highest rate recorded by other state-of-the-art systems. Our network has the capability to increase accuracy while maintaining real-time efficiency in mobile devices. We calculated the operating speed on Pixel 3 (Snapdragon 845) at 22.8 fps. The source code of this research is available on https://github.com/hajizadeh/MobileDenseNet . Highlights: Increase accuracy while maintaining speed and real-time efficiency. A new network by the name of "MobileDenseNet" suitable for embedded systems. A light neck "FCPNLite" for mobile devices that aid with small object detection.