Real-Time Classification of Chicken Parts in the Packaging Process Using Object Detection Models Based on Deep Learning

Processes, Journal Year: 2025, Volume and Issue: 13(4), P. 1005 - 1005

Published: March 27, 2025

Chicken meat plays an important role in the healthy diets of many people and has a large global trade volume. In chicken sector, some production processes, traditional methods are used. Traditional part sorting often manual time-consuming, especially during packaging process. This study aimed to identify classify parts for their input process with highest possible accuracy speed. For this purpose, deep-learning-based object detection models were An image dataset was developed classification by collecting data different parts, such as legs, breasts, shanks, wings, drumsticks. The trained You Only Look Once version 8 (YOLOv8) algorithm variants Real-Time Detection Transformer (RT-DETR) variants. Then, they evaluated compared based on precision, recall, F1-Score, mean average precision (mAP), Mean Inference Time per frame (MITF) metrics. Based obtained results, YOLOv8s model outperformed other YOLOv8 versions RT-DETR obtaining values 0.9969, 0.9950, 0.9807 F1-score, [email protected], [email protected]:0.95, respectively. It been proven suitable real-time applications MITF value 10.3 ms/image.

Language: Английский

---

MDD-DETR: Lightweight Detection Algorithm for Printed Circuit Board Minor Defects

Electronics, Journal Year: 2024, Volume and Issue: 13(22), P. 4453 - 4453

Published: Nov. 13, 2024

PCBs (printed circuit boards) are the core components of modern electronic devices, and inspecting them for defects will have a direct impact on performance, reliability cost product. However, performance current detection algorithms in identifying minor PCB (e.g., mouse bite spur) still requires improvement. This paper presents MDD-DETR algorithm detecting PCBs. The backbone network, MDDNet, is used to efficiently extract features while significantly reducing number parameters. Simultaneously, HiLo attention mechanism captures both high- low-frequency features, transmitting broader range gradient information neck. Additionally, proposed SOEP neck network effectively fuses scale particularly those rich small targets, INM-IoU loss function optimization enables more effective distinction between background, further improving accuracy. Experimental results PCB_DATASET show that achieves 99.3% mAP, outperforming RT-DETR by 2.0% parameters 32.3%, thus addressing challenges defects.

Language: Английский

---

Assessing semantic consistency of pavement markings and signs using street view images – a case study on lane-turning information

International Journal of Digital Earth, Journal Year: 2024, Volume and Issue: 17(1)

Published: Dec. 4, 2024

Road asset management (RAM) is crucial in road construction and maintenance. Previous efforts have focused on the digitization of physical state facilities, such as location condition. However, semantic information conveyed by these instructions, controls, warnings, consistency across multiple facilities has been neglected. Inconsistent can confuse users, disrupt traffic, endanger lives. To address this critical problem, study proposes concept 'semantic space' for presents a comprehensive framework that combines street view images with deep learning techniques to detect, localize, analyze space, specifically focusing lane-turning information. validate effectiveness our framework, we conducted experiments 81 km urban roads Nanjing, Jiangsu, China. The experimental results show method an overall precision 77.6% recall 94.2% detecting defined inconsistency errors. While focuses information, proposed space detection assessment shows promise analyzing inconsistencies other diverse discrete contributing enhanced RAM.

Language: Английский

---

A Comparative Study of Deep Learning Frameworks Applied to Coffee Plant Detection from Close-Range UAS-RGB Imagery in Costa Rica

Remote Sensing, Journal Year: 2024, Volume and Issue: 16(24), P. 4617 - 4617

Published: Dec. 10, 2024

Introducing artificial intelligence techniques in agriculture offers new opportunities for improving crop management, such as coffee plantations, which constitute a complex agroforestry environment. This paper presents comparative study of three deep learning frameworks: Deep Forest, RT-DETR, and Yolov9, customized plant detection trained from images with high spatial resolution (cm/pix). Each frame had dimensions 640 × pixels acquired passive RGB sensors onboard UAS (Unmanned Aerial Systems) system. The image set was structured consolidated UAS-RGB imagery acquisition six locations along the Central Valley, Costa Rica, through automated photogrammetric missions. It evidenced that RT-DETR Yolov9 frameworks allowed adequate generalization mAP50 values higher than 90% mAP5095 54%, scenarios application data augmentation techniques. Forest also achieved good metrics, but noticeably lower when compared to other frameworks. were able generalize detect plants unseen include forest structures within tropical Systems (AFS).

Language: Английский

---