Application Of Visual Saliency Algorithm In Intelligent Industrial Robot

With the continuous development of industrial intelligence,especially the penetration of intelligent industrial robots into all walks of life,the robots are made to work through vision understanding the surrounding work environment is becoming increasingly important.In this thesis,the application of visual saliency algorithms to intelligent industrial robots is investigated in the field of visual saliency principle and CPU/GPU heterogeneous computation in two parts aims to implement a highperformance system for the height analysis of cement mortar.The main research in this thesis is divided into four parts.The first focuses on the implementation principle of the Saliency Filters(SF)algorithm,which proposes the SF inverse parameter optimization method.This optimization method introduces a task-driven model into the SF algorithm,the core idea of which is to find a gradient descent algorithm to make the overall loss function minimal set of optimal parameters.It is demonstrated that by learning from the task data,the reverse parameter optimization method can effectively improve the accuracy of the SF algorithm in a given scenario.Secondly,the time performance of traditional SF algorithm is analyzed in detail,and the time consumption distribution data of each module are given.After that,we analyze in detail the characteristics of each module's memory access behavior and time complexity,and propose the corresponding heterogeneous optimization ideas.Since the optimization effect of heterogeneous computing is highly dependent on the hardware architecture,this thesis gives the various possible time consumption distributions for each module.Optimization ideas,the optimal combination of which is determined by subsequent experimental findings.Based on the above performance analysis,this thesis implements a high-performance parallel SF algorithm in the OpenCL heterogeneous computing framework.Again,compares the time performance of the parallel SF algorithm and the traditional SF algorithm in detail,and the time performance of the parallel SF algorithm is improved.The quantitative analysis of the causes,performance bottlenecks,and further improvement space is given,which also answers the question of the optimal optimization of each module.combinatorial problem.This thesis uses two different sets of test platforms,which on the one hand can show the relationship between heterogeneous computing acceleration ratios and hardware architectures,and on the other hand also provides a good basis for the actual hardware selection provides quantitative criteria.From the test conclusion,each high-performance module in this thesis has 5-20 times faster acceleration compared to the traditional SF algorithm.The temporal performance of the SLIC super pixel clustering module outperforms the previous optimization results.The overall performance of the parallel SF algorithm achieves a 20 times acceleration relative to the conventional SF algorithm with an acceptable loss of accuracy effect,which is of great practical value.This thesis concludes with the implementation of a complete height analysis system for cement mortar,focusing on the implementation details of the height marking module,and gives the Information on the effectiveness and accuracy of the system's application.From the final test conclusions,the system applies the principle of visual saliency to the height analysis of cement mortar,and the system accuracy is comparable to that of the manual identification error is within 10 pixels,achieving the research goal of this thesis.