| China is a major producer of fish products,but there is an imbalance in the development of fish processing processes and production equipment,and a large gap between China and developed countries.It is of great significance to overcome the vulnerable spot in the deep processing of fish products in China for achieving intelligent processing of fish products of great significance.At present,the image segmentation and recognition processing of fish are the research hotspot in recent years.In this paper,it conducted segmentation and identification studies on the section images of Conger Eel of different sizes.(Ⅰ)In the segmentation of section images of fish,a system for acquisition of section images different sizes of Conger Eel was established,in order to contrast enhancement and noise reduction re-processing of the section images of fish.In the aspect of image segmentation,the combined algorithm of SLIC super-pixel segmentation and NJW spectral clustering were used for object segmentation.In the SLIC algorithm,the region connectivity of generated superpixels is enhanced to make the boundary extraction more accurate;in the spectral clustering algorithm,the number of input clusters is judged in advance by SSE,The overall time-consuming speed of the algorithm and the accuracy of fish slice target segmentation are improved.The segmentation effect is analyzed by precision-recall of boundary and segmentation accuracy as evaluation index,to complete accurate extraction of segmentation targets and provide guarantee for subsequent recognition.(Ⅱ)In the recognition of section images of fish,a machine vision recognition system is constructed by pre-processing to enhance images and binary segmentation of images.The color components of R,G and B were extracted from the enhanced denoising images,and the L component of the CLELAB space was extracted from binary segmentation of images.Whereafter,the mean values of the color components of different categories of images were obtained,then to establish the recognition model of the threshold interval of fish section.In the experimental training,50 section images of different categories of large,medium and small each,were used to set up a test set for image recognition based on 20 images of each category;Through the establishment of recognition test data sets for images of different sizes and categories,the recognition verification was carried out based on the YOLOV3 algorithm in deep learning.In terms of improved YOLOV3,First of all,it aims at the problem of high resolution of fish slice image,the input resolution of the original YOLOV3 algorithm were upgraded to meet the high resolution of fish section images,then YOLOV3 feature extraction of the network is improved,in order to change the predicted number to apply the single target detection in fish section images and finally combining multi-scale prediction network and space transformation,the spatial transformation capability of the original algorithm is enhanced.The experimental results showed that in this paper,the precision-recall of boundary reached to 95.48%and the segmentation accuracy rate reached to 90.27%in the segmentation of section images of fish by SLIC super-pixel segmentation and spectral clustering segmentation algorithm.The two evaluation index were higher than other comparative segmentation algorithms,achieving improved the accuracy of segmentation target boundary accuracy and time efficiency of segmentation;in the subsequent image recognition,the YOLOV3 algorithm was verified by comparing it with other network model recognition methods,and the algorithm achieved 89.46%precision-recall and 94.17%accuracy rate in the test set.The effective identification of images of different types of Conger Eel slices was achieved,laying the foundation for subsequent studies of deep processing of the fish body. |
