Lactating sow image segmentation method integrated with FCN and threshold segmentation

Abstract

The invention discloses a lactating sow image segmentation method integrated with FCN and threshold segmentation. The method comprises the steps: collecting a video image of a sow, and building a sow segmentation video image library; building an FCN sow segmentation model, segmenting a test image through the model, and obtaining an FCN sow image segmentation result; building a bounding minimum-area rectangular frame according to the FCN segmentation result, carrying out the Otsu threshold segmentation of the gray scale image and H component of the region, and obtaining a threshold segmentation result; carrying out the fusion of the FCN segmentation result and the threshold segmentation result, and obtaining a final segmentation result of a sow image. On the basis of FCN, the method integrates with the multi-channel Otsu threshold segmentation technology, can effectively complement for a local regional loss while the FCN segmentation effect is not reduced, and improves the segmentation accuracy.

Description

technical field [0001] The present invention relates to the technical field of image segmentation, and more particularly, relates to a method for image segmentation of lactating sows that integrates a fully convolutional network (FCN) network and multi-channel threshold segmentation. Background technique [0002] The health and maternal behavior of lactating sows are related to the economic benefits of the entire pig farm, so it is particularly important to monitor the behavior of lactating sows. The traditional sow condition monitoring method is to observe the daily behaviors of sows such as exercise, eating, and lactation manually for a long time, and judge the sow's physical condition and maternal behavior based on experience, so as to take further relevant measures. This method is not only time-consuming and labor-intensive, but also easily leads to misjudgment. Using computer vision technology to automatically monitor sow behavior is a better choice to replace manual m...

AI Technical Summary

Problems solved by technology

In 2014, Zhu Weixing of Nanjing Agricultural University and others used the mixed Gaussian to update the background model, combined with the maximum information entropy threshold, but this method is not suitable for long-term stationary or slow-moving target prospects

In 2015, the team used the maximum information entropy threshold to perform secondary segmentation on group-raised piglets to obtain the target foreground of the pigs. However, the effect of this method was not good when the target foreground and background were not significantly different.

FCN can better avoid problems such as uneven illumination, random noise, and image distortion through multi-layer convolution, pooling, etc., and has made great breakthroughs in the field of image segmentation. Applied research is almost blank

Since FCN uses simple bilinear interpolation for upsampling, it is easy to lose local information of the image and generate holes

In addition, in the case of insufficient samples or a single sample, FCN shows poor generalization ability and is prone to under-segmentation.

Images