Multi-view Images Research Articles

Enhancing endoscopic scene reconstruction with color-aware inverse rendering through neural SDF and radiance fields.

Virtual surgical training is crucial for enhancing minimally invasive surgical skills. Traditional geometric reconstruction methods based on medical CT/MRI images often fall short in providing color information, which is typically generated through pseudo-coloring or artistic rendering. To simultaneously reconstruct both the geometric shape and appearance information of organs, we propose a novel organ model reconstruction network called Endoscope-NeSRF. This network jointly leverages neural radiance fields and Signed Distance Function (SDF) to reconstruct a textured geometric model of the organ of interest from multi-view photometric images acquired by an endoscope. The prior knowledge of the inverse correlation between the distance from the light source to the object and the radiance improves the real physical properties of the organ. The dilated mask further refines the appearance and geometry at the organ's edges. We also proposed a highlight adaptive optimization strategy to remove highlights caused by the light source during the acquisition process, thereby preventing the reconstruction results in areas previously affected by highlights from turning white. Finally, the real-time realistic rendering of the organ model is achieved by combining the inverse rendering and Bidirectional Reflectance Distribution Function (BRDF) rendering methods. Experimental results show that our method closely matches the Instant-NGP method in appearance reconstruction, outperforming other state-of-the-art methods, and stands as the superior method in terms of geometric reconstruction. Our method obtained a detailed geometric model and realistic appearance, providing a realistic visual sense for virtual surgical simulation, which is important for medical training.

A fast scene geometric modeling approach for digital twins combining neural rendering and model retrieval

ABSTRACT Digital twin technology is a crucial driving technology for the realization of Industry 4.0, which enables simulation, analysis, and prediction by constructing geometric scenes that correspond to physical scenes. Current automation in scene modeling is limited, relying heavily on costly and inefficient manual operations. This hinders digital twin technology progress. This paper proposes a framework for a fast scene geometric modeling method that combines neural rendering and model retrieval for digital twins. Neural rendering techniques first train the collected image data, and the point cloud data of physical entities are rendered. Then based on the semantic mapping algorithm between the point cloud data of physical entities and the multi-view image data of 3D CAD models proposed in this study, the corresponding geometric models are retrieved from the geometric asset library by inputting the point cloud data, and all the retrieved geometric models are embedded in the geometric scene to complete the scene geometric modeling. Finally, in a case study of the digital twin-based scene construction for decommissioned lithium battery dismantling, the effectiveness of this method is demonstrated, which can improve the speed and reduce the cost of scene geometric modeling.

SCA-PVNet: Self-and-cross attention based aggregation of point cloud and multi-view for 3D object retrieval

To address 3D object retrieval, substantial efforts have been made to generate highly discriminative descriptors for 3D objects represented by a single modality, such as voxels, point clouds, or multiview images. It is promising to leverage complementary information from multimodal representations of 3D objects to further improve retrieval performance. However, multimodal 3D object retrieval has rarely been developed or analyzed for large-scale datasets. In this paper, we propose a self-and-cross-attention-based aggregation of point clouds and multiview images (SCA-PVNet) for 3D object retrieval. With deep features extracted from point clouds and multi-view images, we design two types of feature aggregation modules, namely the in-modality aggregation module (IMAM) and the cross-modality aggregation module (CMAM ), for effective feature fusion. IMAM leverages a self-attention mechanism to aggregate multiview features, whereas CMAM exploits a cross-attention mechanism to interact with point-cloud and multiview features. The final descriptor of a 3D object for object retrieval can be obtained by concatenating the aggregated feature outputs of both modules. Extensive experiments and analyses were conducted on four datasets ranging from small to large scales to demonstrate the superiority of the proposed SCA-PVNet over state-of-the-art methods. In addition to achieving state-of-the-art retrieval performance, our method is more robust in challenging scenarios when views or points are missing during inference.

A Robust Registration Method for Multi-view SAR Images based on Best Buddy Similarity

Abstract. Due to the influence of imaging angle and terrain undulation, multi-view synthetic aperture radar (SAR) images are difficult to be directly registered by traditional methods. Although feature matching solves the issue of image rotation and maintains scale invariance, these methods often lead to non-uniformity of interest points and may not achieve subpixel accuracy. The traditional template matching method makes it difficult to generate correct matches for multi-view SAR oblique images. In this paper, a multi-view SAR image template matching method based on Best Buddy Similarity (BBS) is proposed to solve the traditional methods' problem. Firstly, the initial correspondences between images are established according to the Range-Doppler model of SAR images. Secondly, a sliding window search is performed on the established correspondence, the BBS is calculated, and the subpixel locations of the peaks on the similarity map are estimated to achieve a fine match. In the calculation process of BBS, the SAR-ROEWA operator is used to suppress the speckle noise of SAR images. The experiment demonstrated that SAR-BBS can accurately match SAR images with large rotation angle. The peak value on the search window is significant. The registration accuracy of SAR-BBS outperforms the other state-of-the-art methods.

Automatic Localization of Soybean Seedlings Based on Crop Signaling and Multi-View Imaging.

Soybean is grown worldwide for its high protein and oil content. Weeds compete fiercely for resources, which affects soybean yields. Because of the progressive enhancement of weed resistance to herbicides and the quickly increasing cost of manual weeding, mechanical weed control is becoming the preferred method of weed control. Mechanical weed control finds it difficult to remove intra-row weeds due to the lack of rapid and precise weed/soybean detection and location technology. Rhodamine B (Rh-B) is a systemic crop compound that can be absorbed by soybeans which fluoresces under a specific excitation light. The purpose of this study is to combine systemic crop compounds and computer vision technology for the identification and localization of soybeans in the field. The fluorescence distribution properties of systemic crop compounds in soybeans and their effects on plant growth were explored. The fluorescence was mainly concentrated in soybean cotyledons treated with Rh-B. After a comparison of soybean seedlings treated with nine groups of rhodamine B solutions at different concentrations ranging from 0 to 1440 ppm, the soybeans treated with 180 ppm Rh-B for 24 h received the recommended dosage, resulting in significant fluorescence that did not affect crop growth. Increasing the Rh-B solutions reduced crop biomass, while prolonged treatment times reduced seed germination. The fluorescence produced lasted for 20 days, ensuring a stable signal in the early stages of growth. Additionally, a precise inter-row soybean plant location system based on a fluorescence imaging system with a 96.7% identification accuracy, determined on 300 datasets, was proposed. This article further confirms the potential of crop signaling technology to assist machines in achieving crop identification and localization in the field.

Stower-13: A Multi-View Inspection Image Dataset for the Automatic Classification and Naming of Tension Towers

In high-voltage transmission lines, the tension tower needs to withstand the tension load of the overhead power line for a long time, which is prone to damage, and it is an important part of the inspection in the circuit inspection. In the modern circuit inspection process, operation and maintenance personnel mostly use unmanned aerial vehicles (UAVs) to photograph various parts of the tension tower, obtain inspection images, and manually classify and name the massive inspection images, which is low in accuracy and efficiency. Based on the above problems, this paper collects a large number of real-life UAV inspection images of various parts of a tension tower, and proposes the Stower-13 inspection image dataset, which is used to train the classification model to achieve automatic classification and naming of inspection images. Based on this dataset, this paper also proposes an improved MobileViT model, in which the Scale-Correlated Pyramid Convolution Block Attention Block (SCPCbam) module is introduced, which adds the Convolution Block Attention Module (CBAM) to the four branches of the original Scale-Correlated Pyramid Convolution (SCPC) module, so as to strengthen the ability of image multi-scale information extraction and improve the classification accuracy. This paper discusses a number of experiments on the model, and the experimental results show that the dataset proposed in this paper helps the model to understand the feature information. At the same time, the improved MobileViT model has a strong ability to extract image spatial feature information, the classification accuracy is higher than that of other models of the same type, so it is able to cope with a wide range of problems that arise in the course of practice, and it meets the practical needs of automatically naming transmission line inspection images.

Real-time volume rendering with octree-based implicit surface representation

Recent breakthroughs in neural radiance fields have significantly advanced the field of novel view synthesis and 3D reconstruction from multi-view images. However, the prevalent neural volume rendering techniques often suffer from long rendering time and require extensive network training. To address these limitations, recent initiatives have explored explicit voxel representations of scenes to expedite training. Yet, they often fall short in delivering accurate geometric reconstructions due to a lack of effective 3D representation. In this paper, we propose an octree-based approach for the reconstruction of implicit surfaces from multi-view images. Leveraging an explicit, network-free data structure, our method substantially increases rendering speed, achieving real-time performance. Moreover, our reconstruction technique yields surfaces with quality comparable to state-of-the-art network-based learning methods. The source code and data can be downloaded from https://github.com/LaoChui999/Octree-VolSDF.

Towards multi-views cloud retrieval accounting for the 3-D structure collected by directional polarization camera

The structure of a cloud is three-dimensional (3-D). Nevertheless, cloud research is predominantly developed in one-dimensional (1-D) hypothesis from passive optical satellite measurements. Cloud retrieval algorithms typically do not consider the parallax displacement of clouds in multi-view images. Inversion of cloud properties based on two-dimensional (2-D) coordinates cannot solve multi-view radiation changes which are caused by macro-structure of a cloud in 3-D space. In this study, we included the 3-D structure of clouds by repositioning cloud pixels within a 3-D coordinate system utilizing the directional polarization camera (DPC) data before conducting cloud property retrieval. It is followed by identification of location of cloud in each observational view’s image plane. Observational paths of clouds are rebuilt at each observational view. Furthermore, computer graphics methods are employed to establish spatial relationship among different cloud observational paths. The radiational and geometrical information are relocated in unified 3-D sub-pixel space. Finally, 3-D cloud radiation information could be applied to retrieve cloud properties. The case results demonstrate that our method can effectively reduce mixed pixels of clouds and clear sky by 5.79% while decrease indistinguishable pixels in cloud phase classification by 2.11%. It leads to an enhancement in the classification accuracy of water clouds and ice clouds by 3.30% and 4.03%, respectively. Therefore, it is concluded that considering the impact of cloud structure before retrieval would significantly improve accuracy.

Abstract PO1-07-12: Inverse modeling with surface temperature accurately detects the presence of breast cancer

Abstract Breast cancer affects over 250,000 women in the United States every year. Patient outcomes including overall survival significantly decline with large sized tumors and nodal involvement on initial presentation. The introduction of mammogram screening has been instrumental in the early detection of cancer. However, it has lower sensitivity in dense breast tissue and limited specificity which leads to unnecessary additional invasive testing. Several methods such as magnetic resonance imaging (MRI), computed tomography (CT) and ultrasound imaging are used as adjunctive methods. However, there is still a need for cost-effective techniques without additional radiation exposure for breast cancer screening for use with mammography. We propose an adjunctive screening method using surface temperature measurements and inverse breast modeling for early detection of breath cancer. Methods: We enrolled patients who presented with mammogram detected breast tumors that were confirmed to be cancerous lesions by biopsy. We used prone position steady state infrared imaging of bilateral breasts to measure surface temperatures and multi-view MRI images to create patient specific 3D breast models. The inverse technique applies Levenberg-Marquard algorithms (LMA) and utilizes commercial software for thermal transport modeling. The inverse technique predicts a heat generation map normalized for baseline surface temperatures to localize the presence of an underlying tumor. Results: A total of 25 breast tumors (diameter range from 5 to 27 mm) from 24 patients were included in the analysis with a median patient age of 67 years. One patient had bilateral breast tumors. Mammogram data showed inclusion of mixed breast tissue composition including dense and extremely dense breast tissue. LMA accurately detected the tumor in all 25 patients with maximum absolute errors of 7 mm in location of the tumor in the breast and 1.78 mm in diameter. The accuracy of detection was not affected by tumor histology which included invasive ductal and lobular carcinomas, ductal and lobular carcinoma in-situ and atypical ductal hyperplasia. No heat sensitive tumors were detected in the 23 contralateral breasts which acted as internal negative controls. Conclusion: Infrared temperature profiles and inverse modeling successfully detected malignant breast tumors with no missed tumors or false positive results and could be used as adjunctive screening along with mammography, especially in patients with dense breast tissue. Citation Format: Nithya Sritharan, Carlos Gutierrez, Isaac Perez-Raya, Jose-Luis Gonzalez-Hernandez, Alyssa Owens, Donnette Dabydeen, Lori Medeiros, Satish Kandlikar, Pradyumna Phatak. Inverse modeling with surface temperature accurately detects the presence of breast cancer [abstract]. In: Proceedings of the 2023 San Antonio Breast Cancer Symposium; 2023 Dec 5-9; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2024;84(9 Suppl):Abstract nr PO1-07-12.

Computational 3D topographic microscopy from terabytes of data per sample

We present a large-scale computational 3D topographic microscope that enables 6-gigapixel profilometric 3D imaging at micron-scale resolution across >110 cm2 areas over multi-millimeter axial ranges. Our computational microscope, termed STARCAM (Scanning Topographic All-in-focus Reconstruction with a Computational Array Microscope), features a parallelized, 54-camera architecture with 3-axis translation to capture, for each sample of interest, a multi-dimensional, 2.1-terabyte (TB) dataset, consisting of a total of 224,640 9.4-megapixel images. We developed a self-supervised neural network-based algorithm for 3D reconstruction and stitching that jointly estimates an all-in-focus photometric composite and 3D height map across the entire field of view, using multi-view stereo information and image sharpness as a focal metric. The memory-efficient, compressed differentiable representation offered by the neural network effectively enables joint participation of the entire multi-TB dataset during the reconstruction process. Validation experiments on gauge blocks demonstrate a profilometric precision and accuracy of 10 µm or better. To demonstrate the broad utility of our new computational microscope, we applied STARCAM to a variety of decimeter-scale objects, with applications ranging from cultural heritage to industrial inspection.

IMVS: Integrating multi-view information on multiple scales for 3D object recognition

3D object recognition is a fundamental task in 3D computer vision. View-based methods have received considerable attention due to their high efficiency and superior performance. To better capture the long-range dependencies among multi-view images, Transformer has recently been introduced into view-based 3D object recognition and achieved excellent performance. However, the information among views on multiple scales is not utilized sufficiently in the existing Transformer-based methods. To address this limitation, we proposed a 3D object recognition method named iMVS to integrate Multi-View information on multiple Scales. Specifically, for the single-view image/features at each scale, we adopt a hybrid feature extraction module consisting of CNN and Transformer to jointly capture local and non-local information. For the extracted multi-view image features at each scale, we develop a feature transfer module including a view Transformer block to achieve the information transfer across views. Following a sequential process of the single-view feature extraction and multi-view feature transfer on multiple scales, the multi-view information is sufficiently interacted. Subsequently, the multi-scale features with multi-view information are fed into our designed feature aggregation module to generate a category-specific descriptor, where the adopted channel Transformer block facilitates the descriptor to be more expressive. Coupling with these designs, our method can fully exploit the information embedded within multi-view images. Experimental results on ModelNet40, ModelNet10 and a real-world dataset MVP-N demonstrate the superior performance of our method.

Fast 3D Reconstruction for Unknown Space Targets via Neural Radiance Field

In this study, a novel residual space object 3D model reconstruction approach is proposed. First, several low-cost CubeSats are assigned to carry out multi-view image capture in the vicinity of the space target, yielding a series of images with relative pose information. Then, a fast 3D reconstruction approach based on NeRF is introduced, in which real-time performance and reconstruction quality are improved for fast on-board 3D reconstruction. Finally, a simulation test is designed to verify the performance of the proposed method, results show that the proposed NeRF-based method is superior to traditional method in terms of reconstruction speed and texture richness.

Deep-learning based 3D birefringence image generation using 2D multi-view holographic images

Refractive index stands as an inherent characteristic of a material, allowing non-invasive exploration of the three-dimensional (3D) interior of the material. Certain materials with different refractive indices produce a birefringence phenomenon in which incident light is split into two polarization components when it passes through the materials. Representative birefringent materials appear in calcite crystals, liquid crystals (LCs), biological tissues, silk fibers, polymer films, etc. If the internal 3D shape of these materials can be visually expressed through a non-invasive method, it can greatly contribute to the semiconductor, display industry, optical components and devices, and biomedical diagnosis. This paper introduces a novel approach employing deep learning to generate 3D birefringence images using multi-viewed holographic interference images. First, we acquired a set of multi-viewed holographic interference pattern images and a 3D volume image of birefringence directly from a polarizing DTT (dielectric tensor tomography)-based microscope system about each LC droplet sample. The proposed model was trained to generate the 3D volume images of birefringence using the two-dimensional (2D) interference pattern image set. Performance evaluations were conducted against the ground truth images obtained directly from the DTT microscopy. Visualization techniques were applied to describe the refractive index distribution in the generated 3D images of birefringence. The results show the proposed method’s efficiency in generating the 3D refractive index distribution from multi-viewed holographic interference images, presenting a novel data-driven alternative to traditional methods from the DTT devices.

RadarTCN: Lightweight Online Classification Network for Automotive Radar Targets Based on TCN.

Automotive radar is one of the key sensors for intelligent driving. Radar image sequences contain abundant spatial and temporal information, enabling target classification. For existing radar spatiotemporal classifiers, multi-view radar images are usually employed to enhance the information of the target and 3D convolution is employed for spatiotemporal feature extraction. These models consume significant hardware resources and are not applicable to real-time applications. In this paper, RadarTCN, a novel lightweight network, is proposed that achieves high-accuracy online target classification using single-view radar image sequences only. In RadarTCN, 2D convolution and 3D-TCN are employed to extract spatiotemporal features sequentially. To reduce data dimensionality and computational complexity, a multi-layer max pooling down-sampling method is designed in a 2D convolution module. Meanwhile, the 3D-TCN module is improved through residual pruning and causal convolution is introduced for leveraging the performance of online target classification. The experimental results demonstrate that RadarTCN can achieve high-precision online target recognition for both range-angle and range-Doppler map sequences. Compared to the reference models on the CARRADA dataset, RadarTCN exhibits better classification performance, with fewer parameters and lower computational complexity.

Text-image conditioned diffusion for consistent text-to-3D generation

By lifting the pre-trained 2D diffusion models into Neural Radiance Fields (NeRFs), text-to-3D generation methods have made great progress. Many state-of-the-art approaches usually apply score distillation sampling (SDS) to optimize the NeRF representations, which supervises the NeRF optimization with pre-trained text-conditioned 2D diffusion models such as Imagen. However, the supervision signal provided by such pre-trained diffusion models only depends on text prompts and does not constrain the multi-view consistency. To inject cross-view consistency into diffusion priors, some recent works finetune the 2D diffusion model via multi-view data, but still lack fine-grained view coherence. To tackle this challenge, we incorporate multi-view image conditions into the supervision signal of NeRF optimization, which explicitly enforces fine-grained view consistency. With such stronger supervision, our proposed text-to-3D method effectively mitigates the generation of floaters (due to excessive densities) and completely empty spaces (due to insufficient densities). Our quantitative evaluations on the T3Bench dataset demonstrate that our method achieves state-of-the-art performance over existing text-to-3D methods. We will make the code publicly available.

3D medical image segmentation based on semi-supervised learning using deep co-training

In recent years, artificial intelligence has been applied to 3D COVID-19 medical image diagnosis, which reduces detection costs and missed diagnosis rates with higher predictive accuracy, and diagnostic efficiency. However, the limited size and low quality of clinical 3D medical image samples have hindered the segmentation performance of 3D models. Therefore, we propose a 3D medical image segmentation model based on semi-supervised learning using co-training. Multi-view and multi-modal images are generated using spatial flipping and windowing techniques to enhance the spatial diversity of 3D image samples. A pseudo label generation module based on confidence-weights is employed to generate reliable pseudo labels for non-annotated data, thereby increasing the sample size and reducing overfitting. The proposed approach utilizes a three-stage training process: firstly, training a single network based on annotated data; secondly, incorporating non-annotated data to train a dual-modal network and generate pseudo labels; finally, jointly training six models in three dimensions using both annotated and pseudo labels generated from multi-view and multi-modal images, aiming to enhance segmentation accuracy and generalization performance. Additionally, a consistency regularization loss is applied to reduce noises and accelerate convergence of the training. Moreover, a heatmap visualization method is employed to focus on the attention of features at each stage of training, providing effective reference for clinical diagnosis. Experiments were conducted on an open dataset of 3D COVID-19 CT samples and a non-annotated dataset from TCIA, including 771 NIFTI-format CT images from 661 COVID-19 patients. The results of 5-fold cross-validation show that the proposed model achieves a segmentation accuracy of Dice=73.30 %, ASD=10.633, Sensitivity=63.00 %, and Specificity=99.60 %. Compared to various typical semi-supervised learning 3D segmentation models, it demonstrates better segmentation accuracy and generalization performance.

A Reflective Object Reconstruction Method Based on Neural Radiance Field

We propose a neural rendering-based 3D reconstruction method for reconstructing the geometry and BRDF of reflective objects from multi-view images captured in unknown environments. Multi-view reconstruction of reflective objects is extremely challenging because specular reflections are view-dependent and therefore violate multi-view consistency, which is the cornerstone of most multi-view reconstruction methods. Recent neural rendering techniques can model the interaction between ambient light and object surfaces to adapt to view-dependent reflections, making it possible to reconstruct reflective objects from multi-view images. However, accurately modeling ambient light in neural rendering is difficult. We propose a two-step approach to solve this problem. First, by introducing view-dependent photometric losses, our method accurately reconstructs the geometry of reflective objects. Then, with the object geometry fixed, we use more accurate sampling to recover the BRDF of the ambient light and object. Experiments show that our method is able to accurately reconstruct the geometry and BRDF of reflective objects from only RGB images without knowing the ambient light and object masks.

An ensemble deep learning diagnostic system for determining Clinical Activity Scores in thyroid-associated ophthalmopathy: integrating multi-view multimodal images from anterior segment slit-lamp photographs and facial images.

Thyroid-associated ophthalmopathy (TAO) is the most prevalent autoimmune orbital condition, significantly impacting patients' appearance and quality of life. Early and accurate identification of active TAO along with timely treatment can enhance prognosis and reduce the occurrence of severe cases. Although the Clinical Activity Score (CAS) serves as an effective assessment system for TAO, it is susceptible to assessor experience bias. This study aimed to develop an ensemble deep learning system that combines anterior segment slit-lamp photographs of patients with facial images to simulate expert assessment of TAO. The study included 156 patients with TAO who underwent detailed diagnosis and treatment at Shanxi Eye Hospital Affiliated to Shanxi Medical University from May 2020 to September 2023. Anterior segment slit-lamp photographs and facial images were used as different modalities and analyzed from multiple perspectives. Two ophthalmologists with more than 10 years of clinical experience independently determined the reference CAS for each image. An ensemble deep learning model based on the residual network was constructed under supervised learning to predict five key inflammatory signs (redness of the eyelids and conjunctiva, and swelling of the eyelids, conjunctiva, and caruncle or plica) associated with TAO, and to integrate these objective signs with two subjective symptoms (spontaneous retrobulbar pain and pain on attempted upward or downward gaze) in order to assess TAO activity. The proposed model achieved 0.906 accuracy, 0.833 specificity, 0.906 precision, 0.906 recall, and 0.906 F1-score in active TAO diagnosis, demonstrating advanced performance in predicting CAS and TAO activity signs compared to conventional single-view unimodal approaches. The integration of multiple views and modalities, encompassing both anterior segment slit-lamp photographs and facial images, significantly improved the prediction accuracy of the model for TAO activity and CAS. The ensemble multi-view multimodal deep learning system developed in this study can more accurately assess the clinical activity of TAO than traditional methods that solely rely on facial images. This innovative approach is intended to enhance the efficiency of TAO activity assessment, providing a novel means for its comprehensive, early, and precise evaluation.

Multiview Type-2 kernelized fuzzy C-means clustering with local information for noisy color image segmentation

Focusing on the currently available multi-view fuzzy clustering algorithms, many of which frequently lack robustness and are hence less frequently used in image segmentation. We present a multi-view fuzzy clustering image segmentation algorithm in this research, along with an autonomous view-weight learning mechanism. Firstly, to ensure that each view has the best view weight, the algorithm adds a view weight factor. Secondly, it introduces the weighted fuzzy factor and the kernel distance metric, the role of the weighted fuzzy factor is to collect the local spatial information and local grey scale information to preserve as much of the image’s detailed information as feasible during segmentation. The role of the kernel distance metric is to lessen the influence of outliers and noisy points on image segmentation. Finally, the technique for resolving the issue of image uncertainty and fuzzy factor selection introduces the concept of interval type-2 fuzzy c-means clustering. Numerous experiments on different images demonstrate that the proposed algorithm in this paper is more robust than previous multi-view fuzzy clustering algorithms for solving noise image segmentation problems. It is also more effective at segmenting images contaminated by noise and can better retain the detailed information in the image.

Improving deep learning based segmentation of scars using multi-view images

The utilization of deep learning for scar segmentation in photographs enables automated and non-contact quantitative analysis of skin scars. Meanwhile, multi-view photographs are commonly employed to capture the 3D information of scars. In this paper, we propose a two-stage deep learning based segmentation framework for delineating scars from surrounding skin, leveraging multi-view images to achieve enhanced segmentation results compared to single-view approaches. In the first stage, a data augmentation method based on 3D reconstruction and view interpolation is proposed. The generated images are used in a semi-supervised setting to train a single-view segmentation network. In the second stage, a multi-view co-segmentation network (MVCSNet) is proposed to exploit the mutual information between views and to further refine the segmentation. The multi-view feature interaction module (MVFI) uses the prior segmentation results from the first stage, computes feature similarities across views, and optimizes the features. The proposed method was evaluated on two multi-view image datasets containing linear scars and patchy scars, respectively. The results show that the proposed data augmentation method can improve the generalization of the model, particularly for the dataset with smaller size. Comparative analyses demonstrate the superior performance of MVCSNet over other deep learning based segmentation or co-segmentation algorithms.