Methodology for evaluating complex object contour detection accuracy in SLIC-based image segmentation DOI Creative Commons
Bohdan Lukashchuk, Y. V. Shabatura

Scientific Bulletin of UNFU, Journal Year: 2024, Volume and Issue: 34(8)

Published: Dec. 23, 2024

This paper investigates the application of Simple Linear Iterative Clustering (SLIC) algorithm for complex object image segmentation, on example images human body injuries. The study solves problem lack quantitative evidence regarding SLIC's performance in high-precision area and boundary assessment lesion a digital with wound injury it. A comprehensive methodology is developed to evaluate efficacy across various resolutions. research utilizes combined dataset 3696 from Foot Ulcer Segmentation Challenge (FUSeg) WoundSeg datasets. Bayesian optimization utilized fine-tune SLIC hyperparameters, focusing number segments compactness. Results indicate that demonstrates consistent different implementations, achieving Dice scores around 0.84 Soft Boundary F1 0.55. reveals optimal can be defined relative spatial dimensions input image, maximal dimension *2 being most effective value. thorough analysis segmentation metrics conducted, including IoU, Score, Score. introduces employs Score – modification novel metric designed provide more nuanced evaluation detection accuracy while offering smoother landscape. proves particularly valuable assessing objects them tasks. Importantly, this presents an idealized SLIC-based approach, where superpixels are optimally using ground-truth masks establish upper bound performance. compared pre-trained FUSeg UNet model, showcasing superior generalization capability diverse types. On dataset, approach achieved score 0.84, significantly outperforming model (0.12 score). As result, provides insights improving methods highlights need further developing superpixel classification real-world scenarios. findings also highlight potential approaches addressing challenges data types limited training data.

Language: Английский

Risk Prediction and Control Study of a Multitower Separation Process Based on DQRA and Bi-LSTM DOI

Guangchao Ji,

Xuejing Li,

Mingzhang Wang

et al.

ACS Sustainable Chemistry & Engineering, Journal Year: 2025, Volume and Issue: unknown

Published: Jan. 13, 2025

Due to the complexity of multitower separation process (the ethylene as an example) and numerous variables, traditional risk analysis methods cannot meet needs modern enterprises. In this paper, intelligent dynamic quantitative assessment combining bidirectional long short-term memory network (DQRA-Bi-LSTM) is proposed for early warning. First, a simulation carried out obtain working condition data. Based on data from actual conditions plant, preliminary performed using Dow Chemical Fire Explosion Index (F&EI) method. Then, are quantitatively converted values definitions predicted by (Bi-LSTM). Finally, genetic algorithms (GAs) introduced control risk. The method applied demethanization system distillation predict values. application results two cases show that able threshold 0.2 h in advance successfully value below threshold.

Language: Английский

Citations

0
Methodology for evaluating complex object contour detection accuracy in SLIC-based image segmentation DOI Creative Commons
Bohdan Lukashchuk, Y. V. Shabatura

Scientific Bulletin of UNFU, Journal Year: 2024, Volume and Issue: 34(8)

Published: Dec. 23, 2024

This paper investigates the application of Simple Linear Iterative Clustering (SLIC) algorithm for complex object image segmentation, on example images human body injuries. The study solves problem lack quantitative evidence regarding SLIC's performance in high-precision area and boundary assessment lesion a digital with wound injury it. A comprehensive methodology is developed to evaluate efficacy across various resolutions. research utilizes combined dataset 3696 from Foot Ulcer Segmentation Challenge (FUSeg) WoundSeg datasets. Bayesian optimization utilized fine-tune SLIC hyperparameters, focusing number segments compactness. Results indicate that demonstrates consistent different implementations, achieving Dice scores around 0.84 Soft Boundary F1 0.55. reveals optimal can be defined relative spatial dimensions input image, maximal dimension *2 being most effective value. thorough analysis segmentation metrics conducted, including IoU, Score, Score. introduces employs Score – modification novel metric designed provide more nuanced evaluation detection accuracy while offering smoother landscape. proves particularly valuable assessing objects them tasks. Importantly, this presents an idealized SLIC-based approach, where superpixels are optimally using ground-truth masks establish upper bound performance. compared pre-trained FUSeg UNet model, showcasing superior generalization capability diverse types. On dataset, approach achieved score 0.84, significantly outperforming model (0.12 score). As result, provides insights improving methods highlights need further developing superpixel classification real-world scenarios. findings also highlight potential approaches addressing challenges data types limited training data.

Language: Английский

Citations

0