Abstract:
Recently, given the new trends to higher efficiency and automation in livestock farming, research of livestock health monitoring through computer vision has been an active area. Our team has concentrated on pig health monitoring for some time. It was found that pig contour segmentation and feature extraction are unstable and disturbed by pig manure and uneven illumination distribution in the rough environment of a pig house. In this paper, an image fusion method based on the nonsubsampled contourlet transform (NSCT) is presented to improve the stability and accuracy of pig contour segmentation. First, the infrared thermal image and the optical image of a pig, which have been registered, are decomposed by NSCT. After that, a group of low frequency sub-band coefficients and multi-directional band-pass sub-band coefficients of each source image could be obtained. Secondly, different fusion rules for low frequency sub-band coefficients and band-pass sub-band coefficients were proposed. For the fusion of low frequency sub-band coefficients, both the factors of average energy and variance of neighbor area were considered to compute a combined value first. Then, weighted values were obtained based on it. The weighted average results of the coefficients of each image were selected as the final low frequency sub-band coefficients of fusion image. For the band-pass sub-band coefficients, the fusion coefficients were selected based on the rule of maximum energy of a neighbor area. Finally, the fusion image was obtained through inverse NSCT. In experiments, a FLIR T250 infrared thermal imager was used to acquire IR thermal image and optical image at Xima animal husbandry corporation in Zhenjiang city, Jiangsu Province. Before fusing, a pair of IR and optical experiment images with resolution of 452×339 were obtained, which are registered by using the method of contour matching of radial line feature points. Then, a group of tests were completed by using different image fusion methods, including IHS, DWT, contourlet transform and the proposed algorithm. The comparative results show that the proposed algorithm gives the better fusion effect, the average gradient value is about 25% and the quality of edge information remained about 23% higher than the other three methods. The contour segmentation results of fusion images by using Otsu method also demonstrate the good performance of the proposed algorithm. Furthermore, to contrast with different fusion rules in NSCT field, another group of tests illustrated the better segmentation result compared with the other three rules. All the experimental results demonstrated that the proposed algorithm improved the stability and accuracy of pig contour segmentation, which provides a basis for the further research of multi-senor image feature extraction for pig health monitoring.