An Intelligent Pig Weights Estimate Method Based on Deep Learning in Sow Stall Environments

To further the application of artificial intelligence techniques in agriculture, this study proposes an approach based on deep neural network to estimate the live weights of pigs in saw stalls. We design a neural network that uses the back of pigs in top-view depth images as the input and outputs the pig weights estimates. The proposed network, which is based on a Faster-RCNN network with an added regressive branch, integrates the pig detection and live weights regressive network into an end-to-end network. It simultaneously performs pig recognition, location and pig weights estimate. Alternating the training method optimises the proposed network. Image simulation using circles with various overlapping areas and radii is used to prove the efficacy of the proposed network. When the overlap area is greater than 30% of the total area, the proposed network is invalid. Real farm experiments were conducted for three months to construct the back of pigs in top-view depth image data set to train the proposed network. The test results not only prove the relationship between size of back area and pig weights, but also verify that the proposed neural network can accurately estimate pig weights. The study will promote the application of intelligent technique in the livestock farming, and provides some references for intelligent weighing researchers.