Watershed Segmentation Method Research Articles

Abstract 854: An artificial intelligence pipeline for imaging mass cytometry data analysis and its application in ovarian cancer prognostic biomarker discovery

Abstract Imaging mass cytometry (IMC) is a powerful platform which enables high-dimensional, single-cell analysis of cell type and state. However, reliable methods used to analyze the IMC data remain to be developed. We seek to build an artificial intelligence (AI)-based analytics pipeline for imaging mass cytometry (IMC) to increase the accuracy of cell segmentation and spatial information extraction and apply the AI pipeline to analyze the IMC data derived from high-grade serous ovarian cancer samples for patient overall survival prediction. Multiplexed spatial analysis of the tumor microenvironment by IMC was performed on 41 formalin fixed paraffin embedded (FFPE) tissue samples obtained from treatment naïve high- grade serous ovarian cancer patients using a panel of 24 metal-tagged antibodies that are specific to tumor, immune and stromal cell markers. IMC data was collected using a a Helios CyTOF instrument equipped with Hyperion Imaging System (Fluidigm). Mask Region-Convolution Neural Network (Mask R-CNN) model was used for cell segmentation. Cell subtypes were derived by iterative phenograph clustering with different marker combinations and different subsets of cells in each iteration. Tumor area was calculated as a thresholded Gaussian blurred image of the density map of the center-of-mass of Keratin positive tumor cells. All cells with center-of-mass located within the tumor region were classified as intratumoral cells. The average cell composition in the nearest neighborhood of each cell type (distance between the center of mass of two cells < 20 μm) in the tumor cell compartment was computed. The results showed that cell segmentation by Mask R-CNN has a higher accuracy than traditional watershed segmentation. Significantly more granzymeB+ CD8+ T cells and CD11b+ Vista+ cells were found in long-term survivors comparing to short-term survivors in both tumor and stromal cell compartments of the tumor microenvironment. Additionally, CD196+, CD45RO+ and CD73+ cell densities in the tumor cell compartment were significantly lower in long-term than short-term survivors. Our results also showed that the mean numbers of CD73+ cells adjacent to Vista- CD4+ T cells, macrophages and B cells were significantly lower in long-term than short-term survivors. The mean number of granzymeB+ CD8+ T cells adjacent to Vista- CD4+ T cells was significantly higher in long-term than short-term survivors. These data demonstrated that the deep learning-based cell segmentation method achieved higher accuracy than the conventional watershed segmentation method. Furthermore, our AI pipeline can automatically extract cell count and cell neighborhood information in tumor. The cell count and neighborhood information could be further employed as features for machine learning to generate predictive biomarkers for ovarian cancer patient survival. Citation Format: Ying Zhu, Jianting Sheng, Sammy Ferri-Borgogno, Tsz-Lun Yeung, Jared K. Burks, Samuel C. Mok, Stephen T. Wong. An artificial intelligence pipeline for imaging mass cytometry data analysis and its application in ovarian cancer prognostic biomarker discovery [abstract]. In: Proceedings of the Annual Meeting of the American Association for Cancer Research 2020; 2020 Apr 27-28 and Jun 22-24. Philadelphia (PA): AACR; Cancer Res 2020;80(16 Suppl):Abstract nr 854.

Semantic segmentation: A modern approach for identifying soil clods in precision farming

Clod identification and determination of aggregate size distribution have undeniable importance in tillage operations, it affects other missions in the field regarding the energy issue. The objective of the study is accurate identification of clods based on a state-of-the-art method entitled semantic segmentation. In this regard, a deep network in the domain of deep learning was used. As the shape of clods on the surface of the ground is often irregular, object detection algorithms are unable to determine the boundary and, they only put a bounding box around the objects. Thus, there is a gap in this regard with deep learning based semantic segmentation filling this gap properly. In order to identify soil clods and compute their various geometrical specifications in different sizes and then to deal with the clods according to required action in variable-rate applicators, VGG16, which is a deep model, was used for implementing the semantic segmentation task. RGB images obtained from a stereo camera were used for feeding the proposed model as stereo cameras are relatively robust to the ambient light conditions and provide high resolution data in real field conditions. The pixels in each image of the dataset were divided into five groups of clod size, based on the equivalent diameter of each clod. These image pixels were labeled for training the network and extracting the required features. Finally, different soil clods were segmented and classified. A classical watershed segmentation was applied and the deep model was also trained in binary setting to evaluate the performance of the deep model and the results of the binary segmentations were compared. The mean accuracy and mean intersection over union (IoU) of the binary semantic segmentation reached 89.09% and 80.50%, respectively. Also, the watershed segmentation yielded 85.17% and 72.01%, for the same metrics. These results indicated that the semantic segmentation used in this study has significant advantages over the conventional watershed segmentation method.

Integration of Multi-Sensor Data to Estimate Plot-Level Stem Volume Using Machine Learning Algorithms–Case Study of Evergreen Conifer Planted Forests in Japan

The development of new methods for estimating precise forest structure parameters is essential for the quantitative evaluation of forest resources. Conventional use of satellite image data, increasing use of terrestrial laser scanning (TLS), and emerging trends in the use of unmanned aerial systems (UASs) highlight the importance of modern technologies in the realm of forest observation. Each technology has different advantages, and this work seeks to incorporate multiple satellite, TLS- and UAS-based remote sensing data sets to improve the ability to estimate forest structure parameters. In this paper, two regression analysis approaches are considered for the estimation: random forest regression (RFR) and support vector regression (SVR). To collect the dependent variable, in situ measurements of individual tree parameters (tree height and diameter at breast height (DBH)) were taken in a Japanese cypress forest using the nondestructive TLS method, which scans the forest to obtain dense and accurate point clouds under the tree canopy. Based on the TLS data, the stem volume was then computed and treated as ground truth information. Topographic and UAS information was then used to calculate various remotely sensed explanatory variables, such as canopy size, canopy cover, and tree height. Canopy cover and canopy shapes were computed via the orthoimages derived from the UAS and watershed segmentation method, respectively. Tree height was computed by combining the digital surface model (DSM) from the UAS and the digital terrain model (DTM) from the TLS data. Topographic variables were computed from the DTM. The backscattering intensity in the satellite imagery was obtained based on L-band (Advanced Land Observing Satellite-2 (ALOS-2) Phased Array type L-band Synthetic Aperture Radar-2 (PALSAR-2)) and C-band (Sentinel-1) synthetic aperture radar (SAR). All satellite (10–25 m resolution), TLS (3.4 mm resolution) and UAS (2.3–4.6 cm resolution) data were then combined, and RFR and SVR were trained; the resulting predictive powers were then compared. The RFR method yielded fitting R2 up to 0.665 and RMSE up to 66.87 m3/ha (rRMSE = 11.95%) depending on the input variables (best result with canopy height, canopy size, canopy cover, and Sentinel-1 data), and the SVR method showed fitting R2 up to 0.519 and RMSE up to 80.12 m3/ha (rRMSE = 12.67%). The RFR outperformed the SVR method, which could delineate the relationship between the variables for better model accuracy. This work has demonstrated that incorporating various remote sensing data to satellite data, especially adding finer resolution data, can provide good estimates of forest parameters at a plot level (10 by 10 m), potentially allowing advancements in precision forestry.

Watershed segmentation of topographical features on freeform surfaces and its application to additively manufactured surfaces

With the advancement of modern enabling technologies, surfaces of high value-added products are changing from traditional simple geometries to complex geometries that feature freeform shapes. The analysis of topographical features on freeform surfaces is challenging because surface shapes are no longer planar and measurement data often require surface description in the form of triangular meshes. An approach is proposed to extend the watershed segmentation from planar surfaces to freeform surfaces. Surface topography is first extracted from measured surfaces (mesh surfaces) using proper mathematical operations, e.g. fitting and filtration. The watershed segmentation based on Maxwell's theory is extended to 3D triangular meshes and applied to the extracted surface topography defined as a scalar function associated with the vertices of the mesh surface. Critical surface points such as peaks, pits and saddles are identified to construct a Pfaltz graph. Ridge lines and course lines are then traced starting from saddles and following the steepest uphill and downhill paths to peaks and pits, respectively. Wolf pruning of the change tree is employed to merge over-segmented regions. The extended watershed segmentation method is applied to segment surface topography of the additively processed scaffold and lattice surfaces.

Compatibility of Enhancement and Segmentation of Digital Image Processing in Medical Applications

The proposed work is an attempt to investigate the compatibility of enhancement and segmentation within the digital image processing framework in the medical field. The correction nonuniform illumination enhancement technique and marker-controlled watershed segmentation method were conducted in order to calculate the red blood cells in digital images of blood samples. Twenty-five patients at Medical City Teaching Hospital in Baghdad participated in this study. The attained outcomes of the two mentioned methods were compared with the real results, which were obtained by a digital camera coupled with a high technology microscope. The comparison illustrated large matching between two proposed techniques and real outcomes throughout very short computation time in spite of limited features of the used computer and available MATLAB environment.

Pore-throat characterization of unconsolidated porous media using watershed-segmentation algorithm

Pore-throat characteristics of porous media play important roles in various flow processes. However, pore size distribution (PSD) and throat size distribution (TSD) of various porous media have rarely been reported. Watershed-segmentation (WS) is a promising pore-throat identification method. To avoid over- and under-segmentations, a trial-and-error method was devised and validated using artificially ordered packings. Afterwards, we analyzed porous media with various characteristics, and comparisons with artificially ordered packing were made. More angular particles generated larger pores and throats. TSD of spherical particles was bimodal, whereas other particles resembled a normal distribution. Pore surface area and volume relation demonstrated a fractal behavior. Compared with medial axis (MA) method, PSD and TSD from WS method were larger but exhibited similar distribution trends. Normalizing the PSD with the median pore size yielded a single Gaussian model, demonstrating a possible PSD model. However, normalization to TSD data does not produce a single Gaussian model.

Identification And Classification Of Brain Tumor Images Using Efficient Classifier

The main objective of this study is to propose a model for finding brain tumor. Failing to detect the tumor in its prior stage will increases the chance of losing a life. So the identification and treatment for the tumor in its prior stages become vital to save the life of a human being. This work uses the Magnetic Resonance Image (MRI) images to identify and classify the benign and malign type brain tumors. Low pass filter is applied to preprocess the MRI image that removes the unwanted background structures at the same time keeps the important portions sharpened. Watershed segmentation method is used for segmenting the tumor affected area independently. The statistical feature extraction method Gray Level Co-occurrence Matrix (GLCM) is applied to take out the imperative features from the segmented tumor. The feature selection is performed using Recursive Feature Elimination- Particle Swarm Optimization (RFE-PSO) method. Ensemble Support Vector Machine (SVM) is applied to classify the tumors into harmless and harmful from the medical image.

Detection of early decayed oranges based on multispectral principal component image combining both bi-dimensional empirical mode decomposition and watershed segmentation method

Detection of early decay caused by fungal infections in citrus fruit still remains one of the major problems in the post-harvest processing and automatic quality grading. A new combination algorithm by merging multispectral principal component image, bi-dimensional empirical mode decomposition and image reconstruction as well as improved watershed segmentation was proposed to detect the early decay in oranges. Segmented principal component analysis based on three wavelength regions including visible and short wavelength near-infrared (500–1050 nm), visible (500–780 nm) and near-infrared (781–1050 nm) was performed to determine the optimal principal component (PC) image that was used to extract the effective wavelength images by weighting coefficient analysis. Seven wavelength images in the spectral region of 500–1050 nm were finally determined to build the multispectral PC images. The bi-dimensional empirical mode decomposition (BEMD) was used to remove noise in the multispectral PC images and further reconstruct images. An improved watershed segmentation method with morphological gradient reconstruction, marker extraction as well as image amendment, was proposed to segment decay regions in fruit by using the reconstructed multispectral PC images. All samples including 220 each of decayed and sound fruit were utilized to assess classification ability of the proposed combination algorithm. The results indicated that identification accuracies of decayed and sound fruit were 97.3% and 100%, respectively. The multispectral principal component image combining both bi-dimensional empirical mode decomposition and watershed segmentation method can be used as an effective tool for detection of early decayed oranges, and it was also promising for development of a fast and low-cost multispectral imaging system.

Detection of Skin Cancer Using SVM, Random Forest and kNN Classifiers.

Most common and deadly type of cancer is Skin cancer. The destructive kind of cancers in skin is Melanoma as well as it can be identified at the initial stage and can be cured completely. For the diagnosis of melanoma, the identification of the melanocytes in the area of epidermis is an essential stage. In this paper the watershed segmentation method is implemented for segmentation. The extracted segments are subjected to feature extraction. The features extracted are shape, ABCD rule and GLCM. The extracted features are then used for classification. The classifiers are kNN (k Nearest Neighbor), Random Forest and SVM (Support Vector Machine). Among different classifiers, the SVM classifier provided better results for the skin lesions classification.

Characterizing Seedling Stands Using Leaf-Off and Leaf-On Photogrammetric Point Clouds and Hyperspectral Imagery Acquired from Unmanned Aerial Vehicle

Seedling stands are mainly inventoried through field measurements, which are typically laborious, expensive and time-consuming due to high tree density and small tree size. In addition, operationally used sparse density airborne laser scanning (ALS) and aerial imagery data are not sufficiently accurate for inventorying seedling stands. The use of unmanned aerial vehicles (UAVs) for forestry applications is currently in high attention and in the midst of quick development and this technology could be used to make seedling stand management more efficient. This study was designed to investigate the use of UAV-based photogrammetric point clouds and hyperspectral imagery for characterizing seedling stands in leaf-off and leaf-on conditions. The focus was in retrieving tree density and the height in young seedling stands in the southern boreal forests of Finland. After creating the canopy height model from photogrammetric point clouds using national digital terrain model based on ALS, the watershed segmentation method was applied to delineate the tree canopy boundary at individual tree level. The segments were then used to extract tree heights and spectral information. Optimal bands for calculating vegetation indices were analysed and used for species classification using the random forest method. Tree density and the mean tree height of the total and spruce trees were then estimated at the plot level. The overall tree density was underestimated by 17.5% and 20.2% in leaf-off and leaf-on conditions with the relative root mean square error (relative RMSE) of 33.5% and 26.8%, respectively. Mean tree height was underestimated by 20.8% and 7.4% (relative RMSE of 23.0% and 11.5%, and RMSE of 0.57 m and 0.29 m) in leaf-off and leaf-on conditions, respectively. The leaf-on data outperformed the leaf-off data in the estimations. The results showed that UAV imagery hold potential for reliably characterizing seedling stands and to be used to supplement or replace the laborious field inventory methods.

The Similarity of The Parts of Iris Texture

In the proposed search, new technique for calculating the similarity among parts related to iris images for the selected person and among iris images will be suggested by dividing iris texture into pre decided number of parts and calculating some features of each one of these parts as a texture then calculating degree of similarity among specific part and all other parts of the same iris and with some irises in database according to some accepted rules. The proposed method divided iris texture into eight parts with (45 degrees) for each part, the features of GLCM algorithm will be calculated for parts of iris image then the distance or similarity between specific and required iris with other irises that will be stored in database will be calculated and finally, the similarity calculated between each part (with 45 degrees) of iris and parts of irises stored in database. The proposed system applied some suitable techniques for localized the pupil of the iris image by using watershed segmentation method by drawing arrows between each pixels and 8-nighbouring pixels according to occurrence of some suitable feature then deleting all arrows from selected pixels if another feature doesn’t available in this pixels, this proposed system used for reducing database size for each registered person depending on extracted features from some parts of another person or from extracted features from one part of iris image.

Classification of very high-resolution remote sensing images by applying a new edge-based marker-controlled watershed segmentation method

Recent advances in remote sensing technology have led to the creation of many sensors with very high spatial resolution, which enable researchers to detect and recognize earth surface features precisely in fine detail. There is an increasing demand for the development of algorithms to extract geometrical and spatial information from these kinds of images as the prerequisite for accurate thematic map generation. Among the existing techniques, segmentation-based methods are of great interest due to their capabilities in extracting spatial information of the image. In this article, a new watershed segmentation approach is proposed where the image edge information is exploited to improve the results. Next to its simplicity, the main characteristic of this newly proposed method is that there is no need to set any effective parameter(s), a common drawback confronted by most state-of-the-art segmentation techniques. The method is first applied to a well-known dataset, and the result of the segmentation phase is evaluated. The segmentation and classification phases of our approach are then both applied to selected Pleiades and WorldView-2 images. The final spatial–spectral classification map is compared to the results obtained through three common segmentation-based classification methods. The result of this comparison indicates that our approach outperforms the others by having 94.94% overall accuracy for a Pleiades image and 89.81% overall accuracy for a WorldView-2 image.

Knee-Cartilage Segmentation and Thickness Measurement from 2D Ultrasound.

Ultrasound (US) could become a standard of care imaging modality for the quantitative assessment of femoral cartilage thickness for the early diagnosis of knee osteoarthritis. However, low contrast, high levels of speckle noise, and various imaging artefacts hinder the analysis of collected data. Accurate, robust, and fully automatic US image-enhancement and cartilage-segmentation methods are needed in order to improve the widespread deployment of this imaging modality for knee-osteoarthritis diagnosis and monitoring. In this work, we propose a method based on local-phase-based image processing for automatic knee-cartilage image enhancement, segmentation, and thickness measurement. A local-phase feature-guided dynamic-programming approach is used for the fully automatic localization of knee-bone surfaces. The localized bone surfaces are used as seed points for automating the seed-guided segmentation of the cartilage. We evaluated the Random Walker (RW), watershed, and graph-cut-based segmentation methods from 200 scans obtained from ten healthy volunteers. Validation against manual expert segmentation achieved a mean dice similarity coefficient of 0.90, 0.86, and 0.84 for the RW, watershed, and graph-cut segmentation methods, respectively. Automatically segmented cartilage regions achieved 0.18 mm localization accuracy compared to manual expert thickness measurement.

Fast and Accurate Feature Extraction-Based Segmentation Framework for Spinal Cord Injury Severity Classification

Detection of spinal cord injury (SCI) is one of the major problems in MRI images to detect the affected portion of spinal cord regions using feature sets. Automatic detection of spinal cord atrophy is complex due to change in structure, size, and white matter. Delineating gray matter and white matter are the essential factors that influence the detection of spinal cord atrophy and its severity. Automatic segmentation and classification are accurate methods for detecting the severity of the SCI. Hierarchical segmentation, partitioning segmentation, graph, and watershed segmentation methods are used to find the SCI segments in static fixed positions. Also, these segmentation models result in a high false positive rate due to over segmentation features and noise in the segmented regions. Furthermore, these classification methods fail to segment and detect the severity level in the affected region due to over segmentation. In order to overcome these issues, a novel segment-based classification model is required to find the severity of the injury and to predict the disease patterns on the over segmented regions and features. In the present model, a hybrid image threshold technique is used to segment the spinal cord regions for non-linear SVM classification approach. Among the traditional feature segmentation-based classification models, the proposed threshold-based non-linear SVM has better accuracy for SCI detection. The results proved that the present model is more efficient than the earlier approaches in terms of true positive rate ( $\text {TP}=0.9783$ ) and accuracy (0.9683).

TPVTrack v1.0: a watershed segmentation and overlap correspondence method for tracking tropopause polar vortices

Abstract. A new free modular software package is described for tracking tropopause polar vortices (TPVs) natively on structured or unstructured grids. Motivated by limitations in spatial characterization and time tracking within existing approaches, TPVTrack mimics the expected dynamics of TPVs to represent their (1) spatial structure, with variable shapes and intensities, and (2) time evolution, with mergers and splits. TPVs are segmented from the gridded flow field into spatial objects as restricted regional watershed basins on the tropopause, described by geometric metrics, associated over time by overlap similarity into major and minor correspondences, and tracked along major correspondences. Simplified segmentation and correspondence test cases illustrate some of the appeal, sensitivities, and limitations of TPVTrack, including effective representation of spatial shape and reduced false positive associations in time. Tracked TPVs in more realistic historical conditions are consistent in bulk with expectations of life cycle and mean structure. Individual tracks are less reliable when discriminating among multiple overlaps. Modifications to track other physical features are possible, with each application requiring evaluation.

Early detection of decay on apples using hyperspectral reflectance imaging combining both principal component analysis and improved watershed segmentation method

In addition to other surface quality attributes such as size, color and shape, during sorting of harvested apple fruit, early detection of decay is important due to its infectiousness and potential food safety issue. However, automatic and fast inspection of fruit for decay still remains a major problem for the industry. The use of hyperspectral imaging technique makes it possible to perform detection process automatically. Three spectral regions including Vis-NIR (400–1000 nm), Vis (400–780 nm) and NIR (781–1000 nm) were performed using principal component analysis (PCA) to determine the more effective spectral region and PC vector for distinguishing between sound and decayed tissues. Based on the selected PC, loadings corresponding to each wavelength were analyzed to extract key wavelength images in raw hyperspectral data for multispectral image processing. Two sets of multispectral PC score images from Vis-NIR and NIR regions, respectively, were established. To avoid over-segmentation of traditional standard watershed segmentation, global threshold and Ostu, a novel improved watershed segmentation algorithm based on morphological filtering and morphological gradient reconstruction as well as marking constraint were proposed to segment decayed spots on apples. All samples including 220 decayed and 220 sound fruit were used to assess performance of the proposed algorithm. The classification results indicated that 99% of the decayed fruit and 100% of sound fruit were accurately identified by proposed algorithm based on PC3 score images obtained from multispectral PCA of four key wavelengths in NIR region, respectively. This study demonstrated that multispectral images coupled with the improved watershed segmentation algorithm could be a potential approach for detection of early decay on apples. However, further algorithm optimization is still required obtain higher detection accuracy of decayed apples due to zero tolerance for this type of fruit from consumers and processing industries.

A method for accurately segmenting images of medicinal plant leaves with complex backgrounds

The paper proposes a relatively accurate method of segmenting the leaf images of medicinal plants with complex backgrounds. The method directly takes photos of leaves on branches for the purpose of machine identification and protects the plant from leaf picking; but it also encounters higher difficulties in segmentation as branches, soil and other leaves appear inevitably in the background. On the basis of introducing user interaction when appropriate, we refer to and extend the OTSU concept to accurately detect adjacent outer points of the leaf corresponding to each manually marked edge point of the leaf. Afterwards, these points are selectively connected in a way that is close to but not going through the leaf edge as practically as possible so as to obtain background markers that are close to the leaf edge but correctly located outside the leaf. Last but not the least, by combining the foreground and background markers, we adopt the marker-based watershed segmentation method to obtain the correct segmentation result. As shown from test results based on the self-built database 1 consisting of 440 leaf images with complex backgrounds of 88 species of plants and the widely applied database 2 consisting of 232 leaf images with complex backgrounds, the method leads to a better result in five indexes like Precision and Recall, and is one upon the reported methods.

Change Detection of Polarimetric SAR Images Based on the Integration of Improved Watershed and MRF Segmentation Approaches

Land cover and land use change detection is one of the most important applications of polarimetric synthetic aperture radar (PolSAR) data. In a pixel-based method, the speckle effects and the lack of consideration for spatial information produce false alarms. Therefore, an object-based image analysis has been used to overcome these weaknesses. This paper proposes a novel hybrid image segmentation algorithm for improving the accuracy of land cover change detection. This method consists of three steps. First, the segmentation of two PolSAR images by integrating segmentation techniques, namely, region-based improved watershed and an edge-based coupled Markov random field (MRF). Second, the selection of the optimal ratios of polarimetric features based on a genetic algorithm and Jefferies–Matusita distance criteria. Third, the binary classification of image objects using the ratio of mean pixel values of the corresponding image objects. Compared with the conventional watershed and multiresolution segmentation methods, the improved watershed reduces the speckle effects in PolSAR images and avoids the oversegmentation problem. The use of coupled MRF detects the most accurate edge positions and improves the hybrid segmentation approach. The proposed change detection method, as applied on unmanned aerial vehicle synthetic aperture full polarimetric images, demonstrates an overall accuracy of 90.14% and a kappa coefficient of 0.80. Compared with the pixel-based change detection method based on a support vector machine classifier for the selected subspace of ratios of polarimetric features, the proposed method improved the overall accuracy by 8.72% and the kappa coefficient by 0.18.

An Improved Watershed Segmentation Method for Electron Microscopy Images

The analysis of nerve tissues is very helpful to the study of the nervous system. With the development and application of high-throughput technology, it becomes more important to realize the automated batch segmentation of nerve tissues in the electron microscopy(EM) image. Aiming at this, we propose an improved watershed segmentation method based on multiple deep-CNN models. Firstly, we use the classical watershed segmentation algorithm to obtain some candidate edge pixels in an original EM image. Secondly, the local regions of these candidate pixels are respectively extracted as the samples to be predicted. Thirdly, all candidate edge pixels are judged by a trained pixel classifier, of which the architecture is a deep convolutional neural network. Then, an initial segmentation result is generated based on the judgements. Finally, a complete segmentation image can be obtained after the morphological processing. The training and testing of model are finished on the ISBI dataset. The application of model is realized by merging the prediction results of three models with different scales. Experiments show that our method can spend less cost to have an approximate segmentation result, compared with other methods. Besides, our method can also show a better performance in the details.

Joint entropy based learning model for image retrieval

As one classic technique of computer vision, image retrieval could retrieve the target images from hundreds of thousands of images effectively. Furthermore, with the rapid development of deep learning, the quality of retrieval is increased obviously. However, under normal conditions, the high-quality retrieval is supported by a large number of learning instances. The large number of learning instances not only need much human source in the process of selection, but also need much computing source in the process of computation. More importantly, for some special categories, it’s difficult to obtain a large number of learning instances.Aiming at the problem above, we proposed one joint entropy based learning model which could reduce the number of learning instances through optimizing the distribution of learning instances. Firstly, the learning instances are pre-selected using improved watershed segmentation method. Then, joint entropy model is used for reducing the possibility of double, useless even mistaken instances existence. After that, a database using a large number of images is built up. Sufficient experiments based on the database show the model’s superiority that our model not only could reduce the number of learning instances but also could keep the accuracy of retrieval.