Application Research Based On Deep Learning

With the development of deep learning,especially in computer vision,more and more fields could be accelerated by deep learning-based methods to improve work efficiency and production quality.In the field of animation,manual drawing has always been the biggest factor that restricts the output speed.Although the animation rendered based on the three-dimensional model can greatly improve the production speed,the changes in the strokes and the fluency of the lines are not able to reach the level of human artists.With the rapid development of integrated circuits,the demand for printed circuit boards has increased.In the field of board component error correction,the inspection of cartridge elements has always been performed by manual visual inspection.This work is tedious and error-prone in the presence of dense elements.An automated element type recognition method needs to be proposed.Essential tremor(ET)is a common disease,and its diagnostic process relies heavily on the subjective judgment of experienced doctors,and the diagnostic results lack the support of quantitative data.Therefore,a method that can easily detect the kinematic data of essential tremor patients and automatically perform an objective rating is very necessary.Based on the research on deep learning,this paper proposes solutions for the above three problems.At the same time,according to the characteristics of some tasks,several new network structures are proposed,and a large amount of data is collected and labeled to train the corresponding network.The research contents of this article are as follows:Aiming at the problem of the 3D rendering,a sketch style transform network based on pix2 pix and a multi-scale discriminator is proposed.The modification of the discriminator can help the generator to produce sketches with more details in hand-drawn style.Since there is no available dataset,the symmetry sketch data generation method based on the Zhang-suen line thinning algorithm is proposed to directly generate the corresponding 3D rendering style sketch from the hand-drawn sketch.A two-stage generative model is also used to implement automatic coloring and controllable coloring of the sketch.Aiming at the problem of recognition of circuit component,a framework of element type recognition based on text detection and text recognition is proposed.The natural scene text detection algorithm based on EAST is used as the front-end network to extract the text bounding box,combined with the CRNN algorithm which is based on a convolutional neural network and recurrent neural network to identify specific text content.A large amount of data is collected and labeled to train the network to adapt them to operation condition,and fancy results have been achieved.Aiming at the clinical diagnosis of essential tremor,an automatic diagnosis and auxiliary diagnosis method based on human posture estimation is proposed.The tremor characteristics of ET patients are calculated from diagnosis videos collected in clinical,and these characteristics are used to estimate CRST rating automatically and generate auxiliary information.To achieve a stable human pose estimation results from videos for the motion trajectory detection,Open Pose is used as the basic framework of the video human pose estimation method.And LSTM module is introduced into Open Pose to utilize the information of context.In order to further improve the accuracy of joint position estimation,joint interest regions were determined based on the pose estimation results and human priors.At the same time,a deep convolutional neural network based on CPM was trained to estimate the position of the joint center in the region of interest to further improve the estimation accuracy.With the extracted motion trajectory from the video,the trajectory signal is band-pass filtered,and then the upper and lower envelope surfaces of the trajectory signal are calculated to calculate the amplitude.Besides,the wavelet transform is used to calculate the tremor frequency.Finally,according to the frequency and amplitude standard of the clinical tremor rating scale(CRST),the patient's condition could be rated automatically.At the same time,the frequency and amplitude are also used to generate auxiliary information to help doctors in clinical diagnosis.The experiments show that the automatic rating results have a strong correlation with the clinical diagnosis results(r = 0.84).At the same time,the clinical diagnosis experiments show that the auxiliary information could help the doctor in many aspects in clinical diagnosis,and could improve the stability of the diagnosis.