Robustness Analysis And Evaluation Of Superpixel's Consistency Of Superpixel Segmentation Algorithms

Superpixels are getting increased academic interest from diverse field ranging from computer vision to image understanding and analysis to multimedia.Superpixel gen-eration means to group local pixels according to their characteristics such as color and location.In simple words,all the pixels of the same superpixel have similar or identical features.Compared to pixels,superpixel contains more local information.Further-more,they can adhere to most of the object's boundaries in the images.It has generated a keen interest in computer vision because it provides a convenient way to calculate image characteristics and reduce the complexity of subsequent image processing tasks.It also increases the efficiency of processing time expressively.Another dynamic area of research is the development of superpixel segmentation algorithms.Several such al-gorithms have been published recently.These algorithms are evaluated using different performance evaluation metrics and data-sets that result in discrepancies in the compar-ison of algorithms.It calls for a point of reference for comparing advanced methods and evaluating their advantages and disadvantages.Most of these algorithms computed their results on natural images free of noise and significantly generated excellent results.Though,no comprehensive research has been conducted to evaluate the robustness of superpixel segmentation algorithms for noise patterns common in natural images.In this study,the problem of robustness analysis of superpixel segmentation algo-rithms for common types of noise has been addressed.In this regard,the study has been divided into two important phases.In the first and primary phase of the study,a performance evaluation of the robustness of eleven(11)recently proposed superpixel segmentation algorithms to different types of noise was conducted and is,subsequently,presented.For this purpose,various degrees of 2D Gaussian blur,additive white Gaus-sian noise,and impulse noise were selected and applied to the set of images.These nois-es made the object boundaries weak or added extra information to it,furthermore,they also reduced the image quality.The Berkeley image database was selected for the eval-uation process,and the images were corrupted with noise.The superpixel segmentation algorithms chosen for the robustness analysis includes Simple Linear Iterative Cluster-ing(SLIC),Voronoi Cells(VCells),Flooding-Based Superpixel Generation(FCCS),Bilateral Geodesic Distance(Bilateral-G),Superpixel Via Geodesic Distance(SSS-G),Manifold Slic(M-SLIC),Turbopixels,Superpixels Extracted Via Energy-Driven Sam-pling(SEEDS),Lazy Random Walk(LRW),Real-Time Superpixel Segmentation by DBSCAN Clustering,and Video Supervoxels Using Partially Absorbing Random Walk-s(PARW)algornthms.The performance evaluation process was conducted qualitatively as well as quantitatively.For the quantitative performance comparnson,four(04)performance evaluation parameters including Achievable Segmentation Accuracy(ASA),Compactness,Under-Segmentation Error(USE),and Boundary Recall(BR)were selected.We also intro-duced two(02)new parameters called the Percentage Performance Degradation(PPi)and Rate of Performance Degradation Analysis(Ri).The results of these parameters provided a complete analysis of the overall performance degradation of each algorithm.They also show the efficiency of these algorithms at different levels of noise.The results obtained in the first phase demonstrated that all algorithms suffered performance degra-dation due to noise.For Gaussian blur,Bilateral-G exhibited optimal results for ASA and USE measures,SLIC yielded optimal compactness,whereas FCCS and DBSCAN persisted optimal for BR.In the case of additive Gaussian and impulse noises,FCCS exhibited optimal results for ASA,USE,and BR,whereas,Bilateral-G remained a close competitor in ASA and USE for Gaussian noise only.Additionally,Turbopixel yield-ed optimal performance for compactness for both types of noise.Therefore,through various algorithms performed well under different noise types,no single algorithm was able to generate optimal solutions for all three types of noise across all the performance standards.In the second phase of the study,a new and rigorous algorithm is proposed that is used to assess the consistency of superpixels for different superpixel segmentation al-gorithms.The proposed algorithm extracts the superpixels that remain unchanged over certain levels of noise by adopting the functionality of Jaccard Similarity Coefficient(JSC).Technically,a new measure of similarity Jaccard is developed.It is appended with different superpixel segmentation algorithms to compare the similarity between sets of superpixels(original and noisy).Various degrees of 2D Gaussian Blur,Impulse Noise and a combination of both were chosen to corrupt the image data-set.The new proposed algorithm produces similarity indices of the superpixels(original and noisy)using Jaccard Similarity(JS).For a Superpixel to be considered consistent,its simi-larity index must meet the predefined criteria of the threshold r of JSC.At this phase,four(04)different superpixel segmentation algorithms are selected for superpixels con-sistency evaluation;these include Bilateral geodesic distance(Bilateral-G),Flooding based superpixels generation(FCCS),superpixels via geodesic distance(SSS-G),and Turbopixels(TP).The experimental results obtained by the new algorithm demonstrat-ed that no single algorithm was able to yield optimal outcomes and to different degrees failed to produce and maintain consistent superpixels at each level of noise.The results obtained from both phases of the research showed that no single algorithm was able to produce optimal results for noisy data across all performance measures.The study extends our theoretical as well as an experimental understanding of Superpixel Segmen-tation Algorithms and provides evidence that to solve real-world problems successfully,more robust superpixel segmentation algorithms need to be developed.