Deep multi-task learning for farm animal monitoring from images

Large-scale applications of camera equipment has reduced the cost of image and video acquisition, making image and video processing technology be widely used in daily life. However, most deep learning tasks often require a large amount of labeled data for training purpose. But it is difficult to obtain labeled data in many practical application scenarios, which limits effective applications of deep learning techniques. Multi-task deep learning is a method of artificially combining multiple deep learning tasks in series according to specific task objectives in order to meet the main task requirements. In the application of monitoring farm animals, usually only videos can be used as input, and it is difficult to obtain sufficient and effective training data. Therefore, it is more suitable to adopt the multi-task learning method for the application. This thesis makes the novel contributions on how to use multi-task learning methods to achieve the monitoring of farm animals. Firstly, to handle with the common issue of sparse training data in farm animal deep learning tasks, a target detection method is proposed. It achieves this goal with a small amount of training data in application scenarios, and the object detection results are further improved by adding a monocular depth estimation sub-task. Secondly, an animal body length measurement method is proposed which is based on multiple sub-tasks: monocular depth estimation, target segmentation and key point detection. Compared with traditional 2D length measurement methods, the experiment results show a promised improvement. Finally, a multi-task based animal face recognition method is proposed. Again, the proposed method uses the combination of multiple sub-tasks, including monocular depth estimation and 3D convolution feature extraction. The proposed methods of object detection and face recognition have been tested and evaluated on public datasets and our own data sets. Compared with other methods, the results show that this methods have higher accuracy and better scene adaptability in the practical application scenes.