Research On The Recognition Method Of Pigs With Aggressive Behavior Based On Visual Saliency

With the gradual transformation of China’s pig industry from the ancient traditional breeding methods to the advanced industrial production,the use of digital information technology and scientific management models to improve the health level and breeding efficiency of live pigs,reducing the production cost has become an indispensable development mode of modern pig industry.In today’s large-scale breeding industry,Computer Vision Technology occupies a pivotal position in the field of information technology.It can real-time monitor the daily life of group pigs and automatically alarm for abnormal behavior.Among them,the aggressive behavior of pigs may cause skin damage,so it is easy to bring viral infections and bacterial infections,reduce the breeding efficiency.In order to detect and locate the pig with aggressive behavior in real-time,this paper proposes an identifying method of pigs with aggressive behavior based on visual salience.People can quickly find the regions of interest and ignore the regions of non-interest when they observe simple or complex scenes.This is called visual selective attention mechanism,and the regions of interest are called visual salience area.Saliency target detection is to select the saliency region by imitating the attention mechanism of human visual selection.In this paper,we use visual saliency detection method to select the pig which has aggressive behavior in the pigsty as the region of interest.Firstly,a proper optical flow method is chosen to extract the optical flow information from the collected surveillance video,at the same time,the optical flow information is used to judge the attacking behavior of the pig in the continuous frame.If not,repeat this step for the next frame.If there is,extract the spatial and temporal saliency feature of the pig’s aggressive behavior.The experimental results show that the proposed method can effectively detect whether the pig is aggressive or not.Then,the aggressive pigs were identified by the feature fusion method.In the process of spatial saliency feature extraction,the image is real-time SuperPixel segmented by based on DBSCAN clustering,and then the global contrast of each Super Pixel’s color value on the whole image is calculated as the salience value of the Pixel,and finally,the spatial saliency map is formed.In the extraction of temporal saliency map,the attack region is selected according to the amplitude of the optical flow,the event region and the area of the pig.Then the velocity of the point in the attack area is calculated with the relative motion of other points to get the initial temporal saliency map.Finally,the initial temporal saliency map is normalized after ignoring the non-attack region to form the final temporal saliency graph.In this paper,the fusion method firstly determines the region most likely to be attacked by using temporal saliency map,and then takes advantage of the continuity of color values in the same superpixel of the spatial saliency map as a supplement.Finally,the spatiotemporal saliency map is fused by the method of normalized addition under multiple conditions to identify the pig with aggressive behavior.Using the method proposed in this paper,the accuracy of identifying attacking pigs is 85.43%.It provides a new idea for identifying pigs with aggressive behavior in a group-raising environment.