Study And On-line Verification For Shrimp Classification Algorithm Based On Machine Vision Technology

Shrimp industry has always been the pillar industry of aquaculture industry in China.With the increasement of people's living standard,it is very normal for prople to pursue high quality of shrimp products.Shrimp classification,means the grading works for the impurities in the harvested fresh shrimp or the flawed shrimp in the clustered cooked shrimp after cooking.There are large differences between normal shrimp and the impurities at the aspect of color consistency and completeness.Traditional grading methods were often performed by manual work,which has the shortage of labor-intensive,time-consuming and expensive labor cost.Therefore,this research tries to develop an rapid automatic shrimp classification algorithm and grading equipment with replacement of human.The traditional method for the grading of different shrimp type is mainly the mechanical roller equipment,however,the system is not be effective according to the specific shrimp dataset in this research,so this restricts their applications to shrimp industry.Machine vision technology as one nondestructive detection method,has the advantages of high precision,strong controllability,high response speed and immune to environment nosis.This technology also has the ability to discriminate the subtle shrimp texture difference,which can satisfy the requirement for detecting the flawed shrimp on-line.Regard the shrimp(Penaeus Orientalis)as the research object,on the basis of comprehensive utilizing multiple field knowledge of sensor technology,spectrum technology,geometric algebra,statistics and pattern recognition,etc.The objective of this research explores the feasibility to classify the shrimp using machine vision technology.Firstly,we construct an experiment equipment of shrimp classification on the basis of machine vision technology,realize the functions of rapidly acquiring shrimp images and speed of conveyor belt optimization,etc.Secondly,extract and optimize the appearance characteristics of different shrimp types.According to the traditional combination rules between the features and classifiers,we improved their combination rules and make it adaptive to morphology characteristic expressions of shrimp body with highest capacity.Thirdly,aiming at the shrimp touching phenomenon on-line,we proposed two segmentation algorithms respectively for pairs of touching cooked shrimp and the clustered fresh shrimp,and generalize the latter segmentation algorithm into other application of flexional agriculture products.Finally,we make on-line testing for the proposed algorithm,including detection,segmentation and classification,to demonstrate the competence of these proposed algorithm used for on-line duties,thereby complete the whole shrimp classification production line.The objective of this study is to explore the feasibility of applying the machine vision technology into grading the shrimp and to construct the machine vision production system on-line.This study provides the actual detection algorithm and theoretical foundation for building a large-scale flat shrimp quality evaluating system.The main research contents and conclusions were listed as follows:(1)On the basis of three blowoff valves and the iteration thought,the best iteration model of shrimp grading was studied.The relationship between the iterations and the total test error rate was discussed.Theoretical proof about the iteration process terminates in which the total training error rate decrease to zero was given.Therefore,based on this,the whole set of optimization iteration algorithm was constructed.Total 625 images including 305 sound shrimp and 320 flawed shrimp or catches with eight different kinds of classes were used to test the effectiveness of the proposed algorithm.Experimental results showed that overall classification accuracy can reach 94.34%.Analyze from the research results,we got to know that incomplete shrimp are the main reason reducing the total accuracy.Therefore,two methods respectively called curvature discriminate method(CDM)and arc discriminate method(ADM)were proposed.Random repitition experiments of total 100 times on 349 sound shrimp and 547 incomplete shrimp were used to test the effectiveness of the proposed two algorithms.Experimental results showed that CDM can achieve average accuracy of 89.78%,while ADM can achieve average accuracy of 93.72%.Moreover,ADM is 292ms faster than CDM at aspect of the average time consumption.Therefore,taken from accuracy and time consumption,ADM is better.(2)To improve shrimp classification efficiency on-line,situation of actuator designed as a single blowoff valve was discussed.We improved the combination ways between features and classifiers from the iteration algorithm.One-time methods of removing all the' impurities were proposed.We construct the relationship between the used numbers of combinations and the final total accuracy to optimize the "elite votes"between features and classifiers.A few advantage votes that contribute the final accuracy would be picked up,while other disadvantage votes that decrease the final accuracy would be deleted.The final label decision made by IMAJ algorithm is to be equivalent to perform the procedure that "minority is subordinate to the majority" in a smaller range from these "elite votes".Total 1370 images including 672 sound shrimp and 698 impurities were used to test the effectiveness of the proposed algorithm.The accuracy of IMAJ(91.53%)was compared with six other kinds of classifier combination schemes proposed by Kittler.The schemes include Sum(89.93%),Product(65.99%),Max(89.76%),Min(80%),Median(88.18%)and Majority(89.2%).Comparison results indicate that the combination classifier based on IMAJ rule is superior.This research lay the basis for multi-channel shrimp automatically grading system in the near future.(3)Aiming at the problem of counting error and grading error for the pairs of touching cooked shrimp,methods of on-line segmentation and recognition algorithm were studied.We make an improvement toward the traditional pruning algorithm and construct the training model combined with watershed algorithm to test the performance of the proposed algorithm.352 single shrimp were used to train the classification model,and 247 touching shrimp was used to test the effectiveness of the proposed algorithm under six touching scenarios.Experiment results indicate that the proposed algorithm can successfully separate the touching shrimp with highest segmentation accuracy of 96.51%,which is superior than the highest segmentation accuracy of traditional algorithm with 82.43%.However,this segmentation algorithm is noneffective to the clustered fresh shrimp because of the small greyscale difference between the edge of shrimp and the background.(4)According to the over-segmentation problem produced by the use of the improved watershed method,the clustered touching shrimp problem was studied.Because the pairs of concavity points can be detected in every two touching shrimp,we combined the two detected concavity points with circle fitting.Then we integrated heuristic learning model and unsupervise learning model together to find the third key point,which was used to do the circle fitting.100 clustered fresh shrimp images were used to construct the training model,while 50 clustered fresh shrimp images were used to test the model.Experimental results revealed that the proposed algorithm achieved a mean accuracy of 92.7%across the clustered shrimp dataset.Other two application examples of flexional agricultural products,such as clustered little green pepper and shrimp meat,were also used to test the effectiveness of the proposed algorithm.Segmentation results demonstrated it can successfully segment the images,which indicates the proposed algorithm has the potential to separate clustered flexional agricultural products.This research lay the basis for tiled shrimp automatically grading system in the near future.(5)We built an on-line shrimp grading system based on machine vision technology,mainly including mechanism structure,feeding system,snapping image system,classification system and image handling system.We also optimized several parameters,such as feeding speed,angle of inclination of feed machine,speed of conveyor belt and time of exposure of CCD camera,etc.Two different shrimp samples status including the fresh shrimp and frozen cooked shrimp were used to test the effectiveness and generalization of the iteration algorithms,IMAJ combination rule and the concave spots locating based touching shrimp segmentation algorithm in the on-line experiment.1)1114 fresh shrimp images including seven kinds of fresh flawed shrimp types were tested on-line,and the experiment showed that average total accuracy can reach 94.19%using the iteration algorithm.After adding the segmentation algorithms into the system,the average total accuracy increased 2.23%,and the maximum average throughput can reach 2.14 shrimp per second under the single channel with single graining.2)893 frozen cooked shrimp images including two kinds of flawed cooked shrimp types were also tested on-line,and the experiment showed that average total accuracy can reach 87%using the iteration algorithm.After adding the segmentation algorithms into the system,the average total accuracy increased 4.67%,and the maximum average throughput can reach 2.37 shrimp per second under the single channel with single graining.3)In the process of on-line testing,compared with the improved majority combination methods(IMAJ),the iteration algorithm can acquire a better accuracy,but at the aspect of throughput(shrimp per second),IMAJ algorithm is better.