Cocoon Image Segmentation Method Based on Fully Convolutional Networks

Abstract

Taking different kinds of cocoons as research objects, a method of cocoon image segmentation based on full convolutional network is proposed. Firstly, the cocoon image is collected and labeled to construct a cocoon image segmentation data set. Using vgg19 as the backbone structure of the cocoon image segmentation model, migration learning improves the convergence effect of the model. Then replaced the fully connected layer as a convolutional layer, and used the jump structure to achieve high and low layer cocoon image feature fusion. Finally, the resolution of the cocoon image is restored by the deconvolution layer, and the segmentation effect is further optimized at the end of the model. From the qualitative and quantitative perspectives, the comparison experiment of the segmentation effect of the method and the traditional method is carried out, and the robustness test of the method is performed. The experimental results show that the segmentation effect of this method is better, the segmentation speed is 1.12 s/sheet, the pixel accuracy is 98.70%, and the mean Intersection over Union is 96.32%. Compared with the threshold segmentation, k-means, and Fcn-8s methods, the pixel accuracy is 47%, 48%, and 1% point higher, respectively, and the mean Intersection over union is 46%, 50%, and 3% points higher. For untrained cocoon image data sets, the pixel accuracy and average cross-over ratio of the model are both over 95%, and the robustness is good. This method can effectively segment the cocoon image and provide research basis for researching sorting automation.