**1. Introduction**

Currently, with the advancement of artificial intelligence technology, industries in various fields, ranging from automobiles to the Internet of Things (IoT), are developing. In these industries, artificial intelligence calculates the input of multiple datasets and converts it into the required output data [1,2]. Various types of sensors are used to receive data, among which camera sensors and methods for processing visual information input are active fields of research [3,4]. Object recognition using visual data as learning data for deep learning is used in various methods and has been researched in a variety of fields [5]. However, these data are difficult to process in real time using a processor that has small amount of memory because of the large amount of image data. In addition, to implement artificial intelligence in daily life a lightweight embedded board must be used. However, lightweight embedded boards are not suitable for large computation loads, as they have small memory and power.

The weight reduction of the object recognition algorithm using a camera sensor has always been an important task to be solved, and research is currently being conducted in various ways [6]. Various methods of processing images have been developed for effective implement of algorithms [7]. However, as with all algorithms, there is a tradeoff relationship between accuracy, speed, cost, and amount of computation. The ASM framework has been studied as an effective method for mining most unlabeled or partially labeled data to enhance object detection. The ASM framework can be used to build effective

**Citation:** Yun, H.; Park, D. Efficient Object Detection Based on Masking Semantic Segmentation Region for Lightweight Embedded Processors. *Sensors* **2022**, *22*, 8890. https:// dx.doi.org/10.3390/s22228890

Academic Editors: Yong Liu and Xingxing Zuo

Received: 14 October 2022 Accepted: 12 November 2022 Published: 17 November 2022

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

CNN detectors that require fewer labeled training instances while achieving promising results [8].

This paper introduces an object recognition algorithm based on deep learning to accurately recognize objects in real time. YOLO (You Look Only Once), a deep learningbased object recognition architecture, is currently the most well-known and efficient object recognition algorithm. However, it is too heavy an architecture to use in real-time on a lightweight embedded board. Therefore, the ROI (Region of Interest) is set in the input data to reduce the amount of image processing. Figure 1 shows the overall operation of the algorithm. The ROI can be set using ENet (Efficient Neural network), a semantic segmentation architecture based on deep learning. The the object of interest can be expressed in a specific color using semantic segmentation. By binarizing this expression, the remaining parts other than the recognized object are removed. Because this architecture only recognizes people, it is useful for removing objects other than people. Running YOLO using masked images as input data reduces computation and can be used on a lightweight embedded board, resulting in improved accuracy.

By dividing image processing into two steps in this way, the efficiency can be maximized, and the accuracy does not change rapidly in various environments. By setting the ROI after filtering using segments in the input image, cases that were not recognized by different variables, such as differentiation from the background, can be excluded, increasing the accuracy. Using two deep learning models allows for implementation with higher accuracy and faster execution time. When the amount of computation is reduced and the algorithm is implemented on a lightweight embedded board, its scope of use can be widened considerably.

**Board (LS1028, LX2160, Window based PC)**

**Figure 1.** Overall structure.
