Research On Supervised Information Form In Laboratory Mouse Pose Estimation

Animal experiments are the most reliable research method for conducting biological research,drug testing and other topics.With modern technological developments it is possible to capture pictures of animals in real time and to perform fine-grained pose estimation.The current mainstream work on pose estimation is based around the human body and has achieved excellent results.However,due to the large differences in pose data characteristics between experimental animals and humans,the form of supervised information and the network structure in human pose estimation is not adequate for the fine-grained animal pose estimation.The aim of this research is to design a rational form of supervisory information and network structure in experimental mouse pose estimation to implement refined joint point localization techniques for experimental mice.The following research was carried out on this.(1)Characteristics and effects of different forms of supervisory information in pose estimation.First,the form of the distribution of supervised information and its properties are presented in mainstream pose estimation.Secondly,the effect of different forms of supervised information on the performance of the model in pose estimation is investigated,and the pose estimation tasks to which different forms of supervised information are applicable are summarized.Finally,the challenges faced in achieving refined pose estimation for mouse are described.(2)Selecting the optimal response range of joints based reinforcement learning.The posture data of mouse have the characteristics of small overall joint size and significant size differences between joints.The transformation of joint points into uniform response ranges as supervised information along the human pose estimation corrupts the real data distribution of joints,further leading to poor model accuracy.To address this challenge,a reinforcement learning-based policy network is proposed to select different response ranges for different joints of the truth-value heatmap.First,modeling and coding of the state and action of the reinforcement learning algorithm.Secondly,designing a reasonable role of reward mechanism as auxiliary supervisory information to improve the ability of the strategy network to select the response range combination.The experiments on both datasets show that the strategy selected by reinforcement learning has better performance.In addition,the limitations of this approach are analyzed based on the strategies selected for reinforcement learning,and directions are established for later work.(3)Mouse refined pose estimation based on multi-scale supervised networks.Because of the small size and high consistency of labeling,mouse joints require more refined pose estimation compared to humans.An easy way is to borrow the human pose estimation network while using a small range response heatmap as supervised information,however,the model is difficult to train adequately at this time,resulting in poor accuracy of point localization.To solve the above challenges,a multi-scale supervised network algorithm is proposed.On the one hand,the response distribution of joints at different scales is added to the loss function as a spatial constraint,which solves the problem that it is difficult to adequately train the model under the small response range heatmap,and improves the localization accuracy of the model.On the other hand,the output of each scale predicts joints as additional complementary features fused with the highest scale features,which improves the discrimination power of the network for joint points.Ablation experiments and comparison experiments on two datasets show that the multi-scale supervised network has high accuracy in the task of fine point refinement localization.(4)Mouse refined pose estimation based on sensitivity awareness loss.In response to the problem that the large response range true-value heatmap cannot achieve finegrained point localization although it has the advantages of fast optimization and robustness.A refined pose estimation algorithm for mouse based on sensitivity awareness loss is proposed.On the one hand,the sensitivity-aware loss gives different loss sensitivities to different pixels,making the loss function more sensitive to the error of pixels in the joint region.Increasing the discrimination ability of the model for the joints.On the other hand,the sensitivity-aware loss gives different loss sensitivities to different stage sub-networks,giving full play to the role of intermediate supervision,so that the multi-stage network obtains more stable spatial features and mapping relationships.Ablation experiments and comparison experiments on both datasets show that the sensitivity-aware loss achieves highly accurate point estimates while maintaining the advantage of fast optimization of large response range heatmaps.