Although related researches on segmenting marbling in beef with machine vision technology have previously been conducted, these methods proposed are so susceptible to environmental conditions that segmentation accuracy cannot be guaranteed. Deep learning as one well-accepted algorithm has been widely used in image processing due to its sound performance, however it currently seldom occurs in the field of meat quality testing. This study is primarily based on the Fully Convolutional Networks designated by U-net to achieve the marbling segmentation of beef. The network is composed of 37 layers in total, including convolutional layers, max pooling layers, up-sampling layers and merge layers. 72 images were initially collected and manually labeled, thenthey were virtually divided into calibration set and validation set in accordance with 3:1. By comparing RGB grayscale image, there are obvious bimodal peaks in G grayscale histogram, which turns out to be suitable for segmentation. The epoch of training is set as 2000, and it can be seen from the loss function that the model trends to be stable after 1000 iterations. In an attempt to evaluate the model after training, K-means and OTSU are also employed to segment the validation set. The result ultimately shows that the segmentation effect of U-net is superior to traditional machine vision methods. And the accuracy of U-net is 99.48%, precision is 94.75%, recall is 99.49% and Fl-score is 96.8% via a series of in-depth analysis. After observing the images with non-ideal result, it can draw a conclusion that U-net still maintains good stability, meanwhile it presents stronger generalization and robustness during the process of segmentation. In this study, the application of intelligent algorithm to thesegmentation of beef marbling will possibly provide a way to facilitate intelligent detection of meat product.
展开▼
