Recognition And Application Of Thenar Palmprint Based On Deep Learning

Breast ultrasound imaging has the characteristics of low cost,safety and reliability.It is an irreplaceable and important technique in breast medical examination and imaging diagnosis.However,the problems inherent in breast ultrasound images such as blurred edge details and unclear texture increase the labor cost of lesion classification and affect the diagnostic efficiency of doctors.Therefore,how to improve the image quality and reduce the difficulty of classification has become one of the important problems in the auxiliary diagnosis of breast ultrasound image processing.Based on deep learning technology,this project proposes a super-resolution reconstruction model for breast ultrasound images to address the low contrast and poor resolution issues of current imaging methods.To address the difficulty of observing breast lesion images and the high cost of manual classification,an improved classification network is proposed.Furthermore,an integrated solution for super-resolution reconstruction and classification of breast ultrasound images is proposed to unify the above issues and achieve reconstruction and classification.The main innovation work of this paper is as follows:(1)To address the problem of blurry edges in breast ultrasound images,a super-resolution reconstruction model called RUAGAN is proposed.This model uses a GAN-based generator with RRDB as the basic block and adds local attention(LA)to extract high-frequency information in the image.In order to provide more real gradient feedback,the model uses U-Net network as the discriminator.The experimental results on BUSI show that the reconstructed image textures are more abundant with 4 times magnification factor,and the peak signal-to-noise ratio,structural similarity and average gradient are up to 31.980,0.974 and 7.725,respectively,which are significantly improved compared with the current method.(2)Fine features in breast ultrasound images are easily diluted by clutter,and the cost of manual classification is relatively high.To address this problem,an improved ResNet network structure is proposed as the classification network for breast ultrasound imaging.Local importance pooling is used to amplify the information of fine-grained features to improve classification accuracy.All indicators on BUSI show significant improvements,with an F1 score of up to 85.5%.(3)Aiming at the registration problem of reconstruction and classification tasks and the computational cost brought by complex tasks,an enhanced adversarial network(ARN)network architecture is proposed.By distilling knowledge from the base network to reduce the number of parameters,and using reinforcement learning strategies based on multi-classification tasks to achieve the adversarial relationship between the reconstruction and classification networks.The experiment on BUSI proves that this scheme can simultaneously complete the multi-task training of reconstruction and classification.By using the mutual guidance of classification and reconstruction results,the results are significantly improved compared with the two independent tasks.