Intelligent Recognition On Embryonic Developmental Process Of Fish Egg Based On Computer Vision

Quality problem of fish eggs is fundamental in the fishery; survival rate of eggs not only directly affects the production of the larvae, but also decides the future of fish production. Knowledge about the population production and development characteristics of fish eggs plays an important role in fisheries management and aquaculture ecology experiment. It is useful not only for estimating the development ability of fertilized eggs, analyzing the effect of environmental factors on the development duration of eggs and hatchability, but also for improving the quality and efficiency of the normal development of eggs. Using the transparent fish eggs for the study, a new approach for counting, recognition and classification of fish eggs based on automated visual inspection was proposed. Computer vision technology, image processing method, statistical learning theory and pattern recognition are main techniques used in this research. The main research works are as follows:(1) Research about automatic counting method of transparent fish eggs. Firstly, for solving the problems about existed reflected noises of the petri dish, a model of extracting region of interest (ROI) based on background subtraction method was presented. Then, for improving the image contrast level and removing other random noises, the ROI of image was processed by bottom-hat transformed, gray morphological opening operation and Gamma correction. In the side of image segmentation, an Otsu’s method was used to realize the initial segmentation of enhanced image, and mathematical morphology operations were suggested to joint the image gaps, fill up the holes and remove away the small noises. For the contiguous eggs, an improved watershed segmentation method based on the analysis of area was developed, while effectively reducing the running time of the algorithm, the phenomenon of over segmentation is greatly reduced. The algorithms were implemented on a personal computer based on image processing system. Experiment results shows:the average correct rates is more than 90% and, the method is reliable, objective, rapid, reproducible, flexible and non-destructive compared with traditional methods.(2) Research about image processing methods of fish eggs microscopic image. Firstly, for solving the problem that eggs target’s brightness is lower than the background in the original gray image, a method based on brightness inverting operation was developed to reverse the brightness of eggs and background in the image. Then, a median filter algorithm was used to remove away the random noises. For the preprocessed gray image, a fast image segmentation method based on Otsu’s method and mathematical morphology was developed to segment the egg objects out of the background, and optimize the segmented result. For the extraction of complete eggs object, based on the Sobel Operator and mathematical morphology, an automatic detecting and extracting algorithm of the complete eggs was proposed. In addition, an improved watershed segmentation method based on the analysis of area and circularity was developed to separate the contiguous eggs.Finally, an improved iterative thresholding method was used to re-segment the result image for extracting the kernel of fish egg. Experiment results shows that the proposed method can recognize and segment all of the eggs correctly, and got the recognization rate of 100%.(3) Feature extraction and selection for fish egg object. Firstly, according to the changes of morphological characteristics of fish eggs during development, with RGB images, gray images and bi-level images of 110 individual objects of fish eggs obtained by automatic image processing module of fish eggs microscopic image,18 color features,22 shape features and 11 texture features were extracted from the result images and then a initial 51 demensional multi-character feature space was constructed. Then, to solve the problems of redundancy and mutual interference existed in the initial feature set, and obtain optimal feature subset to describe the egg object accurately, a genetic algorithm (GA) was developed to select the optimal features from this multi-character feature space efficiently, accuracy of recognition is suggested to be the main parameter of fitness function. Finally,16 features were selected as the final feature subset to be used to recoginition and classification of fish eggs’ development process.(4) Recoginition and classification of fish egg developmental stage. To find a good and efficient classification module for identifying the fish egg developmental stage, three kinds of classification methods were developed:Nearest neighbor method, BP neural network and Support vector machine (SVM), specially, for the multi-class SVM (MSVM), a one-against-all MSVM and one-against-one voting based MSVM were proposed respectively. These four modules of classification were tested with the obtained feature vectors using leave-one-out cross validation. The experiment results indicate that the average running time of these four algorithms is 5.8s,1679.1s,51.2s and 29.2s, and their classification rate is 76.2%,71.36%,59.86% and 88.13% individually. So the conclusion is that the algorithm of one-against-one voting based MSVM is best and can satisfy the accuracy and speed requirement of automatic identifying the fish egg developmental stages.