Shengping Lv , Tairan Liang , Kaibin Zhang , Shixin Jiang , Bin Ouyang , Quanzhou Li , Xiaoqing Li
{"title":"用于工业表面缺陷检测的轻量级分层聚合任务排列网络","authors":"Shengping Lv , Tairan Liang , Kaibin Zhang , Shixin Jiang , Bin Ouyang , Quanzhou Li , Xiaoqing Li","doi":"10.1016/j.eswa.2024.125727","DOIUrl":null,"url":null,"abstract":"<div><div>Industrial surface defect detection is crucial for maintaining product quality, but it faces challenges such as complex background interference, numerous small defects, and significant variations in defect characteristics. To address these challenges, this paper introduces a novel lightweight hierarchical aggregation task alignment network (LHATA-Net) designed to enhance detection accuracy, computational efficiency, and generalization. LHATA-Net includes four innovative features: (1) a fast-efficient layer aggregation network (F-ELAN) for efficient feature extraction; (2) a hierarchical multiscale feature enhancement path aggregation network (HMFE-PAN) to improve detection of small defects in complex backgrounds; (3) a lightweight task aligned head (LTA-Head) to optimize feature interaction between classification and localization; and (4) a slide loss function (Slideloss) that integrates slide weighting function with binary cross entropy with logits loss function to tackle sample imbalance. To better validate the detector, we compile a large-scale dataset, DsPCBSD+, which includes real images of surface defects on printed circuit boards from practical industrial production. Experimental results demonstrate that LHATA-Net, with only 3.5M parameters and 18.4G floating point operations per second, achieves an inference speed of 54.2 frames per second. It also achieves average precision of 79.6%, 70.0%, and 85.8% at an intersection over union threshold of 0.5 on two steel surface defect datasets and the DsPCBSD+ dataset, respectively. It ranks first, second, and third compared to state-of-the-art (SOTA) real-time detectors. The Friedman test confirms that LHATA-Net surpasses SOTA detectors in overall performance, highlighting its superiority in practical engineering applications. The code is available at <span><span>https://github.com/Tarzan-Leung/LHATA-Net</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":50461,"journal":{"name":"Expert Systems with Applications","volume":"263 ","pages":"Article 125727"},"PeriodicalIF":7.5000,"publicationDate":"2024-11-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"A lightweight hierarchical aggregation task alignment network for industrial surface defect detection\",\"authors\":\"Shengping Lv , Tairan Liang , Kaibin Zhang , Shixin Jiang , Bin Ouyang , Quanzhou Li , Xiaoqing Li\",\"doi\":\"10.1016/j.eswa.2024.125727\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"<div><div>Industrial surface defect detection is crucial for maintaining product quality, but it faces challenges such as complex background interference, numerous small defects, and significant variations in defect characteristics. To address these challenges, this paper introduces a novel lightweight hierarchical aggregation task alignment network (LHATA-Net) designed to enhance detection accuracy, computational efficiency, and generalization. LHATA-Net includes four innovative features: (1) a fast-efficient layer aggregation network (F-ELAN) for efficient feature extraction; (2) a hierarchical multiscale feature enhancement path aggregation network (HMFE-PAN) to improve detection of small defects in complex backgrounds; (3) a lightweight task aligned head (LTA-Head) to optimize feature interaction between classification and localization; and (4) a slide loss function (Slideloss) that integrates slide weighting function with binary cross entropy with logits loss function to tackle sample imbalance. To better validate the detector, we compile a large-scale dataset, DsPCBSD+, which includes real images of surface defects on printed circuit boards from practical industrial production. Experimental results demonstrate that LHATA-Net, with only 3.5M parameters and 18.4G floating point operations per second, achieves an inference speed of 54.2 frames per second. It also achieves average precision of 79.6%, 70.0%, and 85.8% at an intersection over union threshold of 0.5 on two steel surface defect datasets and the DsPCBSD+ dataset, respectively. It ranks first, second, and third compared to state-of-the-art (SOTA) real-time detectors. The Friedman test confirms that LHATA-Net surpasses SOTA detectors in overall performance, highlighting its superiority in practical engineering applications. The code is available at <span><span>https://github.com/Tarzan-Leung/LHATA-Net</span><svg><path></path></svg></span>.</div></div>\",\"PeriodicalId\":50461,\"journal\":{\"name\":\"Expert Systems with Applications\",\"volume\":\"263 \",\"pages\":\"Article 125727\"},\"PeriodicalIF\":7.5000,\"publicationDate\":\"2024-11-15\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"Expert Systems with Applications\",\"FirstCategoryId\":\"94\",\"ListUrlMain\":\"https://www.sciencedirect.com/science/article/pii/S0957417424025946\",\"RegionNum\":1,\"RegionCategory\":\"计算机科学\",\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"Q1\",\"JCRName\":\"COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"Expert Systems with Applications","FirstCategoryId":"94","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S0957417424025946","RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE","Score":null,"Total":0}
A lightweight hierarchical aggregation task alignment network for industrial surface defect detection
Industrial surface defect detection is crucial for maintaining product quality, but it faces challenges such as complex background interference, numerous small defects, and significant variations in defect characteristics. To address these challenges, this paper introduces a novel lightweight hierarchical aggregation task alignment network (LHATA-Net) designed to enhance detection accuracy, computational efficiency, and generalization. LHATA-Net includes four innovative features: (1) a fast-efficient layer aggregation network (F-ELAN) for efficient feature extraction; (2) a hierarchical multiscale feature enhancement path aggregation network (HMFE-PAN) to improve detection of small defects in complex backgrounds; (3) a lightweight task aligned head (LTA-Head) to optimize feature interaction between classification and localization; and (4) a slide loss function (Slideloss) that integrates slide weighting function with binary cross entropy with logits loss function to tackle sample imbalance. To better validate the detector, we compile a large-scale dataset, DsPCBSD+, which includes real images of surface defects on printed circuit boards from practical industrial production. Experimental results demonstrate that LHATA-Net, with only 3.5M parameters and 18.4G floating point operations per second, achieves an inference speed of 54.2 frames per second. It also achieves average precision of 79.6%, 70.0%, and 85.8% at an intersection over union threshold of 0.5 on two steel surface defect datasets and the DsPCBSD+ dataset, respectively. It ranks first, second, and third compared to state-of-the-art (SOTA) real-time detectors. The Friedman test confirms that LHATA-Net surpasses SOTA detectors in overall performance, highlighting its superiority in practical engineering applications. The code is available at https://github.com/Tarzan-Leung/LHATA-Net.
期刊介绍:
Expert Systems With Applications is an international journal dedicated to the exchange of information on expert and intelligent systems used globally in industry, government, and universities. The journal emphasizes original papers covering the design, development, testing, implementation, and management of these systems, offering practical guidelines. It spans various sectors such as finance, engineering, marketing, law, project management, information management, medicine, and more. The journal also welcomes papers on multi-agent systems, knowledge management, neural networks, knowledge discovery, data mining, and other related areas, excluding applications to military/defense systems.