Deep Learning Semantic Segmentation Research Articles

ABANet: Attention boundary‐aware network for image segmentation

AbstractDeep learning techniques have attained substantial progress in various face‐related tasks, such as face recognition, face inpainting, and facial expression recognition. To prevent infection or the spread of the virus, wearing of masks in public places has been mandated following the COVID‐19 epidemic, which has led to face occlusion and posed significant challenges for face recognition systems. Most prominent masked face recognition solutions rely on mask segmentation tasks. Therefore, segmentation can be used to mitigate the negative impacts of wearing a mask and improve recognition accuracy. Mask region segmentation suffers from two main problems: there is no standard type of masks that people wear, they come in different colours and designs, and there is no publicly available masked face dataset with appropriate ground truth for the mask region. In order to address these issues, we propose an encoder–decoder framework that utilizes a boundary‐aware attention network combined with a new hybrid loss to provide a map, patch, and pixel‐level supervision. We also introduce a dataset called MFSD, with 11,601 images and 12,758 masked faces for masked face segmentation. Furthermore, we compare the performance of different cutting‐edge deep learning semantic segmentation models on the presented dataset. Experimental results on the MSFD dataset reveal that the suggested approach outperforms state‐of‐the‐art, algorithms with 97.623% accuracy, 93.814% IoU, and 96.817% F1‐score rate. Our dataset of masked faces with mask region labels and source code will be available online.

Deep Learning based Semantic Segmentation for Buildings Detection from Remote Sensing Images

Building extraction from remote sensing images is the process of automatically identifying and extracting the boundaries of buildings from high-resolution aerial or satellite images. The extracted building footprints can be used for a variety of applications, such as urban planning, disaster management, city development, land management, environmental monitoring, and 3D modeling. The results of building extraction from remote sensing images depend on several factors, such as the quality and resolution of the image and the choice of algorithm.The process of building extraction from remote sensing images typically involves a series of steps, including image pre-processing, feature extraction, and classification. Building extraction from remote sensing images can be challenging due to factors such as varying building sizes and shapes, shadows, and occlusions. However, recent advances in deep learning and computer vision techniques have led to significant improvements in the accuracy and efficiency of building extraction methods. This research presents a deep learning semantic segmentation architecture-based model for developing building detection from high resolution remote sensing images. The open-source Massachusetts dataset is used to train the suggested UNet architecture. The model is optimized using the RMSProp algorithm with a learning rate of 0.0001 for 100 epochs. After 1.52 hours of training on Google Colab the model achieved an 83.55% F1 score, which indicates strong precision and recall.

SEM Image Processing Assisted by Deep Learning to Quantify Mesoporous γ-Alumina Spatial Heterogeneity and Its Predicted Impact on Mass Transfer.

The pore network architecture of porous heterogeneous catalyst supports has a significant effect on the kinetics of mass transfer occurring within them. Therefore, characterizing and understanding structure-transport relationships is essential to guide new designs of heterogeneous catalysts with higher activity and selectivity and superior resistance to deactivation. This study combines classical characterization via N2 adsorption and desorption and mercury porosimetry with advanced scanning electron microscopy (SEM) imaging and processing approaches to quantify the spatial heterogeneity of γ-alumina (γ-Al2O3), a catalyst support of great industrial relevance. Based on this, a model is proposed for the spatial organization of γ-Al2O3, containing alumina inclusions of different porosities with respect to the alumina matrix. Using original, advanced SEM image analysis techniques, including deep learning semantic segmentation and porosity measurement under gray-level calibration, the inclusion volume fraction and interphase porosity difference were identified and quantified as the key parameters that served as input for effective tortuosity factor predictions using effective medium theory (EMT)-based models. For the studied aluminas, spatial porosity heterogeneity impact on the effective tortuosity factor was found to be negligible, yet it was proven to become significant for an inclusion content of at least 30% and an interphase porosity difference of over 20%. The proposed methodology based on machine-learning-supported image analysis, in conjunction with other analytical techniques, is a general platform that should have a broader impact on porous materials characterization.

Leveraging Deep Learning for Automated Reconstruction of Indoor Unstructured Elements in Scan-to-BIM

Abstract. Achieving automatic 3D reconstruction for indoor scenes is extremely useful in the field of scene understanding. Building information modeling (BIM) models are essential for lowering project costs, assisting in building planning and renovations, as well as improving building management efficiency. However, nearly all current available scan-to-BIM approaches employ manual or semi-automatic methods. These approaches concentrate solely on significant structured objects, neglecting other unstructured elements such as furniture. The limitation arises from challenges in modeling incomplete point clouds of obstructed objects and capturing indoor scene details. Therefore, this research introduces an innovative and effective reconstruction framework based on deep learning semantic segmentation and model-driven techniques to address these limitations. The proposed framework utilizes wall segment recognition, feature extraction, opening detection, and automatic modeling to reconstruct 3D structured models of point clouds with different room layouts in both Manhattan and non-Manhattan architectures. Moreover, it provides 3D BIM models of actual unstructured elements by detecting objects, completing point clouds, establishing bounding boxes, determining type and orientation, and automatically generating 3D BIM models with a parametric algorithm implemented into the Revit software. We evaluated this framework using publicly available and locally generated point cloud datasets with varying furniture combinations and layout complexity. The results demonstrate the proposed framework's efficiency in reconstructing structured indoor elements, exhibiting completeness and geometric accuracy, and achieving precision and recall values greater than 98%. Furthermore, the generated unstructured 3D BIM models keep essential real-scene characteristics such as geometry, spatial locations, numerical aspects, various shapes, and orientations compared to literature methods.

Hyperparameter tuning for deep learning semantic image segmentation of micro computed tomography scanned fiber-reinforced composites

Image segmentation with deep learning models has significantly improved the accuracy of the pixel-wise labeling of scientific imaging which is critical for many quantitative image analyses. This has been feasible through U-Net and related architecture convolutional neural network models. Although the adoption of these models has been widespread, their training data pool and hyperparameters have been mostly determined by educated guesses through trial and error. In this study, we present observations of how training data volume, data augmentation, and patch size affect deep learning performance within a limited data set. Here we study U-Net model training on four different samples of x-ray CT images of fiber-reinforced composites. Because the training process is not deterministic, we relied on seven-fold replication of each experimental condition to avoid under-sampling and observe model training variance. Unsurprisingly, we find greater training data volume strongly benefits individual models’ final accuracy and learning speed while depressing variance among replicates. Importantly, data augmentation has a profound benefit to model performance, especially in cases with a low abundance of ground truth, and we conclude that high coefficients of data augmentation should be used in scientific imaging semantic segmentation models. Future work to describe and measure image complexity is warranted and likely to ultimately guide researchers on the minimum required training data volume for particular scientific imaging deep learning tasks.

Detection of bladder cancer cells using quantitative interferometric label-free imaging flow cytometry.

Bladder cancer is one of the most common cancers with a high recurrence rate. Patients undergo mandatory yearly scrutinies, including cystoscopies, which makes bladder cancer highly distressing and costly. Here, we aim to develop a non-invasive, label-free method for the detection of bladder cancer cells in urine samples, which is based on interferometric imaging flow cytometry. Eight urothelial carcinoma and one normal urothelial cell lines, along with red and white blood cells, imaged quantitatively without staining by an interferometric phase microscopy module while flowing in a microfluidic chip, and classified by two machine-learning algorithms, based on deep-learning semantic segmentation convolutional neural network and extreme gradient boosting. Furthermore, urine samples obtained from bladder-cancer patients and healthy volunteers were imaged, and classified by the system. We achieved accuracy and area under the curve (AUC) of 99% and 97% for the cell lines on both machine-learning algorithms. For the real urine samples, the accuracy and AUC were 96% and 96% for the deep-learning algorithm and 95% and 93% for the gradient-boosting algorithm, respectively. By combining label-free interferometric imaging flow cytometry with high-end classification algorithms, we achieved high-performance differentiation between healthy and malignant cells. The proposed technique has the potential to supplant cystoscopy in the bladder cancer surveillance and diagnosis space.

Not just More, but more diverse: Green landscapes along urban roads may significantly reduce drivers' psychophysiological fatigue

The impact of roadside greenness on driving fatigue in real urban settings has been insufficiently investigated, presenting a critical knowledge gap for researchers, policymakers, professionals, and the public. In this onsite driving experiment, 34 urban residents completed seven driving tasks on different urban road routes in a randomized order with one-day intervals. A total of 238 tasks were conducted, each lasting an hour, assessing psychophysiological, visual, and muscular fatigue. A cardiovascular activity monitor (BioHarness) continuously measured the driver's heart rate, with lower rates indicating reduced psychophysiological fatigue. Visual and muscular fatigue were self-reported using a Visual Analog Scale questionnaire administered before, at the midpoint, and after completing the driving task. Deep transfer learning semantic segmentation analyzed road landscape characteristics and traffic conditions recorded from the drivers' view. Statistical analysis demonstrated that higher mean and variation in greenness significantly predicted lower psychophysiological fatigue after adjusting for multiple covariates. These results indicate that enhancing both the quantity and diversity of green landscapes along urban roads is vital for reducing driver's psychophysiological fatigue. This study reveals that roadside landscapes in urban settings are not trivial decorations, and they should be considered an essential component of transportation infrastructure.

Semantic segmentation of remote sensing images based on dual‐channel attention mechanism

AbstractDue to the inadequate utilization of data correlation and complementarity in the feature extraction process of multimodal remote sensing images, the paper proposes a deep learning semantic segmentation algorithm based on the Dual Channel Attention Mechanism (DCAM). This algorithm uses U‐Net as the backbone, combining the RGB remote sensing image as one input channel with the Convolutional Block Attention Module to extract colour space features. Simultaneously, it utilizes near‐infrared (NIR) as another input channel with the Self‐Attention Module (SAM) to extract shape space features. Finally, by concatenating the multi‐scale attention features of the RGB remote sensing image channel and the NIR remote sensing image channel, it achieves the correlation and complementarity of contextual features between the two modal remote sensing images. Experimental results on the GID‐15 dataset demonstrate that the DCAM algorithm significantly improves the segmentation accuracy, edge segmentation quality, and object segmentation integrity for various types of targets compared to current mainstream segmentation methods.

A Lightweight Deep Learning Semantic Segmentation Model for Optical-Image-Based Post-Harvest Fruit Ripeness Analysis of Sugar Apples (Annona squamosa)

The sugar apple (Annona squamosa) is valued for its taste, nutritional richness, and versatility, making it suitable for fresh consumption and medicinal use with significant commercial potential. Widely found in the tropical Americas and Asia’s tropical or subtropical regions, it faces challenges in post-harvest ripeness assessment, predominantly reliant on manual inspection, leading to inefficiency and high labor costs. This paper explores the application of computer vision techniques in detecting ripeness levels of harvested sugar apples and proposes an improved deep learning model (ECD-DeepLabv3+) specifically designed for ripeness detection tasks. Firstly, the proposed model adopts a lightweight backbone (MobileNetV2), reducing complexity while maintaining performance through MobileNetV2′s unique design. Secondly, it incorporates the efficient channel attention (ECA) module to enhance focus on the input image and capture crucial feature information. Additionally, a Dense ASPP module is introduced, which enhances the model’s perceptual ability and expands the receptive field by stacking feature maps processed with different dilation rates. Lastly, the proposed model emphasizes the spatial information of sugar apples at different ripeness levels by the coordinate attention (CA) module. Model performance is validated using a self-made dataset of harvested optical images categorized into three ripeness levels. The proposed model (ECD-DeepLabv3+) achieves values of 89.95% for MIoU, 94.58% for MPA, 96.60% for PA, and 94.61% for MF1, respectively. Compared to the original DeepLabv3+, it greatly reduces the number of model parameters (Params) and floating-point operations (Flops) by 89.20% and 69.09%, respectively. Moreover, the proposed method could be directly applied to optical images obtained from the surface of the sugar apple, which provides a potential solution for the detection of post-harvest fruit ripeness.

Intelligent Identification and Quantitative Characterization of Pores in Shale SEM Images Based on Pore-Net Deep-Learning Network Model

Among the various shale reservoir evaluation methods, the scanning electron microscope (SEM) image method is widely used. Its image can intuitively reflect the development stage of a shale reservoir and is often used for the qualitative characterization of shale pores. However, manual image processing is inefficient and cannot quantitatively characterize pores. The semantic segmentation method of deep learning greatly improves the efficiency of image analysis and can calculate the face rate of shale SEM images to achieve quantitative characterization. In this paper, the high-maturity shale of the Longmaxi Formation in the Changning area of Yibin City, Sichuan Province, and the low-maturity shale of Beibu Gulf Basin in China are studied. Based on the Pore-net network model, the intelligent identification and quantitative characterization of pores in shale SEM images are realized. The pore-net model is improved from the U-net deep-learning network model, which improves the ability of the network model to identify pores. The results show that the pore-net model performs better than the U-net model, FCN model, DeepLab V3 + model, and traditional binarization method. The problem of low accuracy of the traditional pore identification method is solved. The porosity of SEM images of high-maturity shale calculated by the pore-net network model is between 12 and 19% deviation from the experimental data. The calculated porosity of the SEM image of the low-maturity shale has a large deviation from the experimental data, which is between 14 and 47%. Compared with the porosity results calculated by other methods, the results calculated by pore-net are closer to the true value, which proves that the porosity calculated by the pore-net network model is reliable. The deep-learning semantic image segmentation method is suitable for pore recognition of shale SEM images. The fully convolutional neural network model is used to train the manually labeled shale SEM images, which can realize the intelligent recognition and quantitative characterization of the pores in the shale SEM images. It provides a certain reference value for the evaluation of shale oil and gas reservoirs and the study of other porous materials.

Exploring Uncertainty-Based Self-Prompt for Test-Time Adaptation Semantic Segmentation in Remote Sensing Images

Test-time adaptation (TTA) has been proven to effectively improve the adaptability of deep learning semantic segmentation models facing continuous changeable scenes. However, most of the existing TTA algorithms lack an explicit exploration of domain gaps, especially those based on visual domain prompts. To address these issues, this paper proposes a self-prompt strategy based on uncertainty, guiding the model to continuously focus on regions with high uncertainty (i.e., regions with a larger domain gap). Specifically, we still use the Mean-Teacher architecture with the predicted entropy from the teacher network serving as the input to the prompt module. The prompt module processes uncertain maps and guides the student network to focus on regions with higher entropy, enabling continuous adaptation to new scenes. This is a self-prompting strategy that requires no prior knowledge and is tested on widely used benchmarks. In terms of the average performance, our method outperformed the baseline algorithm in TTA and continual TTA settings of Cityscapes-to-ACDC by 3.3% and 3.9%, respectively. Our method also outperformed the baseline algorithm by 4.1% and 3.1% on the more difficult Cityscapes-to-(Foggy and Rainy) Cityscapes setting, which also surpasses six other current TTA methods.

ENHANCING URBAN PATHFINDING FOR PEDESTRIANS THROUGH FUSION OF MLS AND HMLS DATA

Abstract. Pedestrian pathfinding is crucial for enhancing pedestrian mobility in urban environments. In this research a method to generate navigation graphs based on Mobile Laser Scanning (MLS) and Handheld Laser Scanning (HMLS) data fusion is developed. The input data comprises a 2-kilometer urban street network that integrates both MLS and HMLS data, effectively mitigating sidewalk occlusions caused mainly by parked vehicles. The proposed method encompasses the following steps: (1) Deep Learning semantic segmentation of urban ground elements and navigation-related obstacles, (2) Static vs. Dynamic object differentiation to replicate conditions of unobstructed passage. (3) Sidewalk enrichment with inclination, preservation status, and width. (4) Physical accessibility estimation between sidewalks and crosswalks incorporating curb information. And (5) navigation graph computation based on the enhanced sidewalk data, crosswalks, with accurate node location and connections. Routes were calculated within a Geographical Information System, and results ensure that pedestrians can navigate in urban environments with precision and efficiency.

Pore characterization was achieved based on the improved U-net deep learning network model and scanning electron microscope images

The SEM image method is commonly used in the qualitative characterization of shale pores. The development of shale micro-reservoir pores can be visually observed through SEM images, but the efficiency of manual image processing is low and subjective. The introduction of deep learning greatly improves the efficiency of pore analysis. In this paper, the argon ion polishing SEM image of Longmaxi Formation shale in southern Sichuan is taken as an example. Intelligent identification and quantitative characterization of pores in shale SEM images are realized by Pore-net network model. Pore-net is based on the U-net network model. The way the model reads the data is changed so that the model does not focus on the region of interest. The number of convolutional layers of the model is increased. The Canny edge extraction algorithm is added. It not only reduces the workload of data set production, but also enhances the ability of network model to identify pores. The results show that the deep learning semantic image segmentation method is suitable for pore recognition of shale SEM images. The full convolutional neural network model is used to train the manually labeled shale SEM images, which can realize the intelligent recognition and quantitative characterization of shale SEM images. Compared with FCN and DeepLab V3+ network model, Pore-net performs better. Only 170 data sets are used to train the model. The Pore-net network model still has a good recognition effect on pores, which solves the problem of low accuracy of traditional pore recognition methods. The deviation between the porosity calculated by the Pore-net network model and the experimental data is between 12% and 19%. Compared with the porosity results calculated by the binarization method and other network models, the results calculated by Pore-net are closer to the real values, which proves that the porosity calculated by the Pore-net network model is reliable.

A Deep Learning Method for Arctic Sea Ice Type Classification Based on Active-Passive Microwave Data

The Arctic-scale classification of sea ice type is important in the fields of macro-monitoring in the Arctic, climate change assessment, and long time-series interannual variability of sea ice. To date, a lack of deep learning-based algorithms for detecting Arctic sea ice types from microwave sensors is comparatively rare. In this study, a sea ice type classification algorithm based on the U-Net-CBAM deep learning semantic segmentation network was developed by carrying out research on sea ice type classification algorithms in the Arctic through the integrated use of ASCAT scatterometer and AMSR2 radiometer data. The results of the study show that the proposed method achieves an impressive average accuracy of 94.1% in extracting sea ice categories using AARI ice maps. The results of the algorithm show a high degree of agreement with the OSI-SAF sea ice type daily products.

Sea ice detection network for icebreakers in polar environments with attention-based deeplabv3+ architecture

Shipborne sea ice detection aboard icebreakers plays a paramount role in polar navigation. The continuous evolution of deep learning semantic segmentation networks has promoted the advancement of sea ice detection tasks. At this stage, there are relatively few studies on shipboard sea ice detection, and the accuracy of polar sea ice detection will be reduced due to problems such as blurred sea fog and indistinct boundaries. In this study, a shipboard sea ice detection dataset is constructed, and a sea ice detection method that combines multi-branch attention feature alignment and multi-scale feature extraction is proposed. The heterogeneous receptive field enhancement atrous spatial convolution pooling pyramid module is designed, and the feature alignment module based on the attention mechanism is constructed, which strengthens the model’s extraction of sea ice features and elevates representation performance. Experimental results underscore the heightened precision of our approach in sea ice detection, to some extent alleviating the issue of missed detections in new ice. It constitutes a positive contribution towards advancing shipborne sea ice detection in polar environments.

Intelligently Quantifying the Entire Irregular Dental Structure

Quantitative analysis of irregular anatomical structures is crucial in oral medicine, but clinicians often typically measure only several representative indicators within the structure as references. Deep learning semantic segmentation offers the potential for entire quantitative analysis. However, challenges persist, including segmentation difficulties due to unclear boundaries and acquiring measurement landmarks for clinical needs in entire quantitative analysis. Taking the palatal alveolar bone as an example, we proposed an artificial intelligence measurement tool for the entire quantitative analysis of irregular dental structures. To expand the applicability, we have included lightweight networks with fewer parameters and lower computational demands. Our approach finally used the lightweight model LU-Net, addressing segmentation challenges caused by unclear boundaries through a compensation module. Additional enamel segmentation was conducted to establish a measurement coordinate system. Ultimately, we presented the entire quantitative information within the structure in a manner that meets clinical needs. The tool achieved excellent segmentation results, manifested by high Dice coefficients (0.934 and 0.949), intersection over union (0.888 and 0.907), and area under the curve (0.943 and 0.949) for palatal alveolar bone and enamel in the test set. In subsequent measurements, the tool visualizes the quantitative information within the target structure by scatter plots. When comparing the measurements against representative indicators, the tool’s measurement results show no statistically significant difference from the ground truth, with small mean absolute error, root mean squared error, and errors interval. Bland-Altman plots and intraclass correlation coefficients indicate the satisfactory agreement compared with manual measurements. We proposed a novel intelligent approach to address the entire quantitative analysis of irregular image structures in the clinical setting. This contributes to enabling clinicians to swiftly and comprehensively grasp structural features, facilitating the design of more personalized treatment plans for different patients, enhancing clinical efficiency and treatment success rates in turn.

Bypassing Pre-processing Method in Alzheimer’s Disease Diagnosing using Deep Learning Instance Segmentation

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder that will cause the memory loss of patient and will progressively lead to loss of bodily function that will eventually lead to death. Therefore, diagnosing AD accurately is critical to provide the patients with suitable treatment to delay the progression of AD as well to facilitate the treatment interventions. Recent studies are more dependent on the Deep Learning Semantic Segmentation method to perform the Alzheimer's Disease diagnosis. However, semantic segmentation will segment every single pixel in the images which will affect the precision of the small targets like hippocampal region in MRI images, even though the overall loss is low enough. Therefore, a Deep Learning Instance Segmentation is introduced into the Alzheimer’s disease diagnosis field without using any pre-processing method. In this research, the Mask R-CNN will be used to localize the hippocampal region to do the segmentation, and then classified it as AD or NC. The dataset UTM_ADNI_RAW will be used in this study. The proposed method applied on UTM_ADNI_RAW shows the high accuracy of 92.67%. These results show that the proposed method to segment the hippocampal region without requiring pre-processing techniques has a good accuracy in classifying AD and NC subjects. In conclusion, the proposed Mask R-CNN generated a good result on segmenting the hippocampal region without requiring any pre-processing techniques.

Aedes Aegypti larvae detection and tracking in turbid videos by semantic segmentation using deep learning

Dengue mosquitoes are the only reason for dengue fever. To effectively combat this disease, it is important to eliminate dengue mosquitoes and their larvae. However, there are currently very few computer-aided models available in scientific literature to prevent the spread of dengue fever. Detecting the larvae stage of the dengue mosquito is particularly important in controlling its population. To address this issue, we propose an automated method that utilizes deep learning for semantic segmentation to detect and track dengue larvae. Our approach incorporates a contrast enhancement approach into the semantic neural network to make the detection more accurate. As there was no dengue larvae dataset available, we develop our own dataset having 50 short videos with different backgrounds and textures. The results show that the proposed model achieves up to 79% F-measure score. In comparison, the DeepLabV3, Resnet achieves up to 77%, and Segnet achieves up to 76% F-measure score on the tested frames. The results show that the proposed model performs well for small object detection and segmentation. The average F-measure score of all the frames also indicates that the proposed model achieves a 76.72% F-measure score while DeepLabV3 achieves a 75.37%, Resnet 75.41%, and Segnet 74.87% F-measure score.

Tool wear segmentation in blanking processes with fully convolutional networks based digital image processing

The extend of tool wear significantly affects blanking processes and has a decisive impact on product quality and productivity. For this reason, numerous scientists have addressed their research to wear monitoring systems in order to identify or even predict critical wear at an early stage. Existing approaches are mainly based on indirect monitoring using time series, which are used to detect critical wear states via thresholds or machine learning models. Nevertheless, differentiation between types of wear phenomena affecting the tool during blanking as well as quantification of worn surfaces is still limited in practice. While time series data provides partial insights into wear occurrence and evolution, direct monitoring techniques utilizing image data offer a more comprehensive perspective and increased robustness when dealing with varying process parameters. However, acquiring and processing this data in real-time is challenging. In particular, high dynamics combined with increasing strokes rates as well as the high dimensionality of image data have so far prevented the development of direct image-based monitoring systems. For this reason, this paper demonstrates how high-resolution images of tools at 600 spm can be captured and subsequently processed using semantic segmentation deep learning algorithms, more precisely Fully Convolutional Networks (FCN). 125,000 images of the tool are taken from successive strokes. Selected images are labeled pixel by pixel according to their wear condition and used to train a FCN (U-Net). The approach presented offers the possibility of spatially monitoring wear in high-speed blanking operations in real-time.

Ice floe segmentation and floe size distribution in airborne and high-resolution optical satellite images: towards an automated labelling deep learning approach

Abstract. Floe size distribution (FSD) has become a parameter of great interest in observations of sea ice because of its importance in affecting climate change, marine ecosystems, and human activities in the polar ocean. A most effective way to monitor FSD in the ice-covered regions is to apply image processing techniques to airborne and satellite remote sensing data, where the segmentation of individual ice floes is a challenge in obtaining FSD from remotely sensed images. In this study, we adopt a deep learning (DL) semantic segmentation network to fast and adaptive implement the task of ice floe instance segmentation. In order to alleviate the costly and time-consuming data annotation problem of model training, classical image processing technique is applied to automatically label ice floes in local-scale marginal ice zone (MIZ). Several state-of-the-art (SoA) semantic segmentation models are then trained on the labelled MIZ dataset and further applied to additional large-scale optical Sentinel-2 images to evaluate their performance in floe instance segmentation and to determine the best model. A post-processing algorithm is also proposed in our work to refine the segmentation. Our approach has been applied to both airborne and high-resolution optical (HRO) satellite images to derive acceptable FSDs at local and global scales.