| With the continuous pursuit of intelligence and automatic,the use of computers is developing rapidly.People are no longer satisfied with only using it to deal with objective and regular problems,but hope that it can be endowed with certain”wisdom”to deal with some subjective and irregular problems.And in this process,a new method,deep learning,was born.It is widely used in various fields,such as image,voice and even more creative fields.In this paper,a total of two topics are studied,the first is the object detection and instance segmentation tasks in the field of computer vision,the second is the use of deep learning method of the automatic composition of a specific style.Firstly,in the second chapter of this paper,we introduce the object detection and instance segmentation tasks,which can be seen everywhere in real life and are the cornerstone of more complex tasks.In this paper,a multi-tasking network is proposed which can handle both tasks simul-taneously.The proposed network is improved based on the network that obtained the first place of instance segmentation and the second place of object detection in the COCO 2017 Challenge.The network mainly con-sists of five parts:feature extraction network,feature augmented path,feature fusion operation and boundary box and mask prediction struc-tures.The feature extraction network adopts the feature pyramid network?FPN?,and its backbone network is Res2Net.The module is internally layered and builds residual-like connections,which can extract multi-scale features with better performance.The specific theory and forward flow are introduced in the third chapter;The proposed network uses feature augmented path to speed up the flow of low-dimensional local information,so as to ensure that the final prediction results are jointly acted by high-dimensional and low-dimensional information;We also use feature fusion operation to fuse feature of different scales to avoid information loss;The final two prediction branches are parallel in structure,and the specific network structure and algorithm are shown in the first section of chapter5.The experimental results are presented in the first section of chapter 6.The second topic of this paper is the automatic composition network using attention mechanism.The datasets used by the proposed network are the Bach chorales which have strict four-part structure and a dignified and simple style.We adopt the overall architecture of Generative Adver-sarial Network?GANs?,and use a seperate GAN system to generate and discriminate each voice part.And we insert an attention mechanism both into the generative model and discriminative model,which allows the net-work to selectively emphasize important information and is more effective for long sequences.Specifically,the attention method used in this paper is the self-attention mechanism,which pays more attention to the character-istics of the sequence itself.The theory of attention mechanism is in the fourth chapter,and the complete network structure and algorithm flow are in the second section of the chapter 5.Finally,the network is tested in two ways:the numerical reharmonize experiment and the subjective random sequence generation experiment,in which the reharmonize experiment is compared with the reproduced BiLSTM-GANs[29]results.The specific ex-perimental setup and experimental results are shown in the second section of chapter 6. |
PDF Full Text Request