基于計(jì)算機(jī)視覺(jué)的魚(yú)卵胚胎發(fā)育過(guò)程智能化識(shí)別方法研究
發(fā)布時(shí)間：2018-07-29 11:15
【摘要】：魚(yú)卵質(zhì)量問(wèn)題是漁業(yè)養(yǎng)殖發(fā)展的根本,魚(yú)卵階段成活率的高低不僅直接影響著幼魚(yú)的生產(chǎn)量,也決定著魚(yú)的未來(lái)總產(chǎn)量。開(kāi)展魚(yú)卵發(fā)育過(guò)程研究,估計(jì)魚(yú)卵生物量,準(zhǔn)確鑒別魚(yú)卵發(fā)育期,對(duì)于估計(jì)受精卵的發(fā)育能力,掌握環(huán)境因子對(duì)魚(yú)卵發(fā)育的影響,以及提高魚(yú)卵發(fā)育質(zhì)量和效率都是至關(guān)重要的。本文以透明魚(yú)卵為研究對(duì)象,利用計(jì)算機(jī)視覺(jué)技術(shù)、圖像處理方法、統(tǒng)計(jì)學(xué)習(xí)理論和模式識(shí)別技術(shù)等進(jìn)行了魚(yú)卵發(fā)育過(guò)程自動(dòng)化視覺(jué)檢測(cè)方法的研究,實(shí)現(xiàn)了魚(yú)卵自動(dòng)化計(jì)數(shù)、魚(yú)卵發(fā)育階段自動(dòng)化識(shí)別與分類等操作。主要工作如下：(1)魚(yú)卵自動(dòng)化計(jì)數(shù)方法研究。針對(duì)圖像存在的對(duì)比度偏低,反射光線噪聲干擾比較嚴(yán)重等問(wèn)題,提出了基于背景差法的魚(yú)卵感興趣區(qū)域提取算法；建立了基于底帽變換和形態(tài)學(xué)灰度開(kāi)運(yùn)算的圖像去噪模型,以及基于伽瑪變換的圖像增強(qiáng)模型。在圖像分割方面,采用Otsu自適應(yīng)閾值法對(duì)增強(qiáng)后的圖像進(jìn)行初始分割,并利用形態(tài)學(xué)處理實(shí)現(xiàn)斷裂縫隙的連接、孔洞的填充,以及小目標(biāo)噪聲的去除等后期處理。針對(duì)結(jié)果圖中存在的粘連魚(yú)卵,提出了基于連通域面積分析的改進(jìn)分水嶺分割方法,在有效降低算法運(yùn)行時(shí)間的同時(shí),大幅度降低了過(guò)分割現(xiàn)象。實(shí)驗(yàn)結(jié)果表明：所提算法的計(jì)數(shù)準(zhǔn)確率達(dá)90%以上,且與傳統(tǒng)方法相比,操作上具有可重復(fù)性、靈活性,以及對(duì)魚(yú)卵的無(wú)破壞性；結(jié)果上具有可靠性強(qiáng)、準(zhǔn)確度高、不受主觀影響等優(yōu)勢(shì)。(2)魚(yú)卵個(gè)體發(fā)育顯微圖像處理方法研究。針對(duì)魚(yú)卵個(gè)體圖像存在的魚(yú)卵目標(biāo)亮度低于背景亮度,有少許隨機(jī)噪聲干擾等問(wèn)題,提出了基于亮度取反運(yùn)算和中值濾波處理的預(yù)處理模型。建立了基于Otsu自適應(yīng)閾值法和數(shù)學(xué)形態(tài)學(xué)處理的魚(yú)卵目標(biāo)與背景分離算法。提出了基于改進(jìn)的Sobel算子檢測(cè)法和形態(tài)學(xué)處理的不完整魚(yú)卵去除算法,并利用基于連通域個(gè)數(shù)和圓形度閾值約束的分水嶺分割方法實(shí)現(xiàn)了粘連魚(yú)卵的有效分割。最后采用改進(jìn)的迭代閾值化方法對(duì)魚(yú)卵目標(biāo)結(jié)果灰度圖執(zhí)行二次分割,實(shí)現(xiàn)了魚(yú)卵內(nèi)核目標(biāo)的準(zhǔn)確提取。對(duì)96幅圖像分割實(shí)驗(yàn)的結(jié)果表明本算法可實(shí)現(xiàn)所有魚(yú)卵圖像的正確識(shí)別和分割,識(shí)別率達(dá)100%。(3)魚(yú)卵個(gè)體目標(biāo)特征提取與選擇方法研究。針對(duì)魚(yú)卵個(gè)體目標(biāo)特征空間存在的多質(zhì)性問(wèn)題,以獲取到的110個(gè)魚(yú)卵個(gè)體目標(biāo)RGB圖、灰度圖和二值圖為數(shù)據(jù)基礎(chǔ),分別提取計(jì)算了魚(yú)卵目標(biāo)的18個(gè)顏色特征、22個(gè)形狀特征和11個(gè)紋理特征,構(gòu)成了51維的初始特征集。針對(duì)特征空間存在的冗余,相互干擾等問(wèn)題,以識(shí)別正確率為適應(yīng)度函數(shù)的主要評(píng)價(jià)參數(shù),提出了基于遺傳算法的魚(yú)卵目標(biāo)特征選擇方法,從51維的多質(zhì)特征空間中優(yōu)選出最具分類能力的16個(gè)特征項(xiàng),實(shí)現(xiàn)了特征空間的有效降維。(4)魚(yú)卵發(fā)育階段識(shí)別分類方法研究。在分析多種分類方法特點(diǎn)的基礎(chǔ)上,分別設(shè)計(jì)實(shí)現(xiàn)了最近鄰分類器、BP神經(jīng)網(wǎng)絡(luò)分類器和支持向量機(jī)(SVM)分類器等多種分類算法,用于進(jìn)行魚(yú)卵發(fā)育階段自動(dòng)化識(shí)別的研究。其中在研究利用SVM分類方法實(shí)現(xiàn)魚(yú)卵發(fā)育階段識(shí)別方面,對(duì)基于1對(duì)多的MSVM算法和基于1對(duì)1投票策略的MSVM算法分別進(jìn)行了算法設(shè)計(jì)與研究。最后采用留一交叉驗(yàn)證法對(duì)設(shè)計(jì)的分類器進(jìn)行了測(cè)試驗(yàn)證,結(jié)果表明：提出的四種分類器的平均運(yùn)行時(shí)間為5.8s、1679.1s、51.2s和29.2s;平均分類正確率為76.2%、71.36%、59.86%和88.13%。因此,研究發(fā)現(xiàn)基于1對(duì)1投票策略的MSVM分類器更適于魚(yú)卵發(fā)育階段的自動(dòng)化識(shí)別分類。
[Abstract]:The quality of fish eggs is the root of the development of fishery culture. The survival rate of the fish egg stage not only directly affects the production of young fish, but also determines the total output of the fish in the future. The influence of development and the improvement of the quality and efficiency of fish egg development are very important. In this paper, the automatic visual inspection method of fish eggs development process is studied by computer vision technology, image processing method, statistical learning theory and pattern recognition technology, and the automatic counting of fish eggs is realized by using the transparent fish eggs as the research object. Automatic identification and classification of the development stage of fish eggs. The main work is as follows: (1) study on automatic counting method of fish eggs. Aiming at the low contrast of the image and the serious interference of the reflected light noise, the algorithm based on background difference method is proposed, based on the bottom cap transformation and morphology. The image denoising model of gray scale operation and the image enhancement model based on gamma transform. In the aspect of image segmentation, the Otsu adaptive threshold method is used for the initial segmentation of the enhanced image, and the connection of cracks, the filling of the holes, the removal of the small target noise and so on are realized by the morphological processing. The improved watershed segmentation method based on the area analysis of connected domain is proposed in this paper, which reduces the over segmentation greatly while effectively reducing the running time of the algorithm. The experimental results show that the counting accuracy of the proposed algorithm is above 90%, and the operation is repeatable and flexible compared with the transmission method. The results have the advantages of strong reliability, high accuracy and no subjective influence. (2) study on the microscopic image processing method for the individual development of fish eggs. A preprocessing model of median filter processing. A separation algorithm based on Otsu adaptive threshold method and mathematical morphological processing is established. An incomplete fish egg removal algorithm based on improved Sobel operator detection and morphological processing is proposed, and a watershed score based on the number of connected domains and the threshold of roundness threshold is used. The method realizes the effective segmentation of the fish eggs. Finally, the improved iterative threshold method is used to perform two segmentation of the target gray map of the fish egg target. The accurate extraction of the core target of the fish eggs is realized. The results of the 96 image segmentation experiments show that the algorithm can realize the correct recognition and segmentation of all the fish eggs and the recognition rate is 100%. (3) study on the feature extraction and selection method of individual target of fish eggs. Aiming at the multiple quality problem of the individual target characteristic space of the fish eggs, the RGB map of 110 individual target of fish eggs, gray map and two value map are obtained as the data basis, and 18 color features, 22 shape features and 11 texture features are extracted and calculated respectively. The 51 dimensional initial feature set, aiming at the problem of redundancy and mutual interference in the feature space, to identify the main evaluation parameters of the correct rate as the fitness function, proposed the method of selecting the target feature of the fish eggs based on the genetic algorithm, and optimized the 16 feature items with the most classification ability from the 51 dimensional multi feature space, and realized the feature space. Effective dimensionality reduction. (4) research on identification and classification of fish egg development stage. On the basis of analyzing the characteristics of various classification methods, several classification algorithms, such as nearest neighbor classifier, BP neural network classifier and support vector machine (SVM) classifier, are designed and implemented for automatic identification of fish eggs development stage. Using the SVM classification method to realize the identification of fish egg development stage, the algorithm based on 1 pairs of MSVM algorithm and the 1 pair 1 voting strategy based on MSVM algorithm is designed and studied respectively. Finally, the left one cross validation method is used to test and verify the designed classifier. The results show that the average running time of the four classifiers is proposed. For 5.8S, 1679.1s, 51.2s and 29.2s, the average classification accuracy was 76.2%, 71.36%, 59.86% and 88.13%.. Therefore, the study found that MSVM classifier based on 1 to 1 voting strategies was more suitable for automatic identification of fish eggs development stage.
【學(xué)位授予單位】：中國(guó)農(nóng)業(yè)大學(xué)
【學(xué)位級(jí)別】：博士
【學(xué)位授予年份】：2016
【分類號(hào)】：S917.4;TP391.41
，
本文編號(hào)：2152512
[Abstract]:The quality of fish eggs is the root of the development of fishery culture. The survival rate of the fish egg stage not only directly affects the production of young fish, but also determines the total output of the fish in the future. The influence of development and the improvement of the quality and efficiency of fish egg development are very important. In this paper, the automatic visual inspection method of fish eggs development process is studied by computer vision technology, image processing method, statistical learning theory and pattern recognition technology, and the automatic counting of fish eggs is realized by using the transparent fish eggs as the research object. Automatic identification and classification of the development stage of fish eggs. The main work is as follows: (1) study on automatic counting method of fish eggs. Aiming at the low contrast of the image and the serious interference of the reflected light noise, the algorithm based on background difference method is proposed, based on the bottom cap transformation and morphology. The image denoising model of gray scale operation and the image enhancement model based on gamma transform. In the aspect of image segmentation, the Otsu adaptive threshold method is used for the initial segmentation of the enhanced image, and the connection of cracks, the filling of the holes, the removal of the small target noise and so on are realized by the morphological processing. The improved watershed segmentation method based on the area analysis of connected domain is proposed in this paper, which reduces the over segmentation greatly while effectively reducing the running time of the algorithm. The experimental results show that the counting accuracy of the proposed algorithm is above 90%, and the operation is repeatable and flexible compared with the transmission method. The results have the advantages of strong reliability, high accuracy and no subjective influence. (2) study on the microscopic image processing method for the individual development of fish eggs. A preprocessing model of median filter processing. A separation algorithm based on Otsu adaptive threshold method and mathematical morphological processing is established. An incomplete fish egg removal algorithm based on improved Sobel operator detection and morphological processing is proposed, and a watershed score based on the number of connected domains and the threshold of roundness threshold is used. The method realizes the effective segmentation of the fish eggs. Finally, the improved iterative threshold method is used to perform two segmentation of the target gray map of the fish egg target. The accurate extraction of the core target of the fish eggs is realized. The results of the 96 image segmentation experiments show that the algorithm can realize the correct recognition and segmentation of all the fish eggs and the recognition rate is 100%. (3) study on the feature extraction and selection method of individual target of fish eggs. Aiming at the multiple quality problem of the individual target characteristic space of the fish eggs, the RGB map of 110 individual target of fish eggs, gray map and two value map are obtained as the data basis, and 18 color features, 22 shape features and 11 texture features are extracted and calculated respectively. The 51 dimensional initial feature set, aiming at the problem of redundancy and mutual interference in the feature space, to identify the main evaluation parameters of the correct rate as the fitness function, proposed the method of selecting the target feature of the fish eggs based on the genetic algorithm, and optimized the 16 feature items with the most classification ability from the 51 dimensional multi feature space, and realized the feature space. Effective dimensionality reduction. (4) research on identification and classification of fish egg development stage. On the basis of analyzing the characteristics of various classification methods, several classification algorithms, such as nearest neighbor classifier, BP neural network classifier and support vector machine (SVM) classifier, are designed and implemented for automatic identification of fish eggs development stage. Using the SVM classification method to realize the identification of fish egg development stage, the algorithm based on 1 pairs of MSVM algorithm and the 1 pair 1 voting strategy based on MSVM algorithm is designed and studied respectively. Finally, the left one cross validation method is used to test and verify the designed classifier. The results show that the average running time of the four classifiers is proposed. For 5.8S, 1679.1s, 51.2s and 29.2s, the average classification accuracy was 76.2%, 71.36%, 59.86% and 88.13%.. Therefore, the study found that MSVM classifier based on 1 to 1 voting strategies was more suitable for automatic identification of fish eggs development stage.
【學(xué)位授予單位】：中國(guó)農(nóng)業(yè)大學(xué)
【學(xué)位級(jí)別】：博士
【學(xué)位授予年份】：2016
【分類號(hào)】：S917.4;TP391.41
，
本文編號(hào)：2152512
本文鏈接：http://www.lk138.cn/shoufeilunwen/xxkjbs/2152512.html
最近更新
教材專著
