Research And Application Of Improved Faster R-CNN In Object Parameter Measurement

At present,large-scale breeding plants with tens of thousands of pigs have certain needs for real-time understanding of target body conditions.The weight parameter is one of the key information reflecting the body condition.But it is time-consuming and labor-intensive to obtain by artificial means.At the same time,it will affect the normal growth and development of the target.To solve this problem,the paper constructs a new network model using deep learning technology.Without affecting the normal life of the measuring object,the network makes the measured weight closer to the true value.Firstly,the paper studies the mainstream algorithm of deep learning technology for target detection—Faster R-CNN.Images in natural scenes often contain cluttered backgrounds.It will affect further analysis of important goals.Extracting major goals from the background is the basis for subsequent research.Blocking is a common problem affecting detection results in real life.The paper mainly studies the detection performance of Faster R-CNN algorithm under different occlusion conditions and analyzes the influence of different occlusion ratios on the detection results.The experimental results show that when the occlusion ratio of the two targets is above 30%,there will be serious missed detection.It will affect the subsequent weight prediction.Secondly,the paper improves on Faster R-CNN algorithm and expands the function of Faster R-CNN algorithm.The improved algorithm solves the problem of target parameter prediction.The paper builds an end-to-end deep neural network capable of parallel processing of three tasks including target recognition,target location and target parameter prediction.The paper uses a unique training method to complete the training of the model.The target recognition and positioning tasks use the joint training method to learn.At the same time,they alternately train with the target parameter prediction task and complete their learning relatively independently.The extended network is combined with the unique training method to obtain network model.The model test results show that the model achieves a 1.55% the average relative error of parameter prediction without occlusion.Under the occlusion condition,the average relative error of the parameter prediction when the training data satisfies the occlusion ratio less than 30% is 1.64%.Finally,the paper uses the improved network to predict the body weight of pig in the limit field under natural scene.We label images by manual and build a dataset.Using this dataset trains the network to obtain a model.Then we test the model and analyze the results.The test results show that the average absolute error of the predicted weight of the network model is 0.644 kg and the average relative error is 0.374%.The network performance is perfect to meet the requirements of practical applications.