Research And Implementation Of Sperm Target Detection Algorithm Based On Multi-scale Perception Fusion

Infertility is increasingly common in modern society,affecting more than 10% of couples worldwide,and at least 30% of these infertility cases are caused by men.Sperm detection plays an important role in the clinical diagnosis of male infertility.With the development of computer vision technology,target detection technology has become more mature.Using computer vision technology to assist sperm detection has become a time-saving and labor-saving standardized technology.Using computer vision algorithms to detect sperm can also reduce the burden on medical staff and ensure accuracy.Moreover,the deployment of mobile terminals can reduce the cost of testing for patients.However,computer vision-based sperm detection faces three problems:(1)The size of sperm in sperm images taken by mobile phones and microscopes may be very small,and small targets have insufficient semantic features and features when using deep learning feature extraction.Blur and other problems;(2)because the sperm target resolution is low,and its shape is simpler than ordinary objects,it is easy to over-fit some key areas during the training process;(3)deep learning training excessive model parameters are not conducive to mobile deployment,and there will be problems such as insufficient video memory during the deployment process.In view of the above problems,the research work of this paper is as follows:(1)Aiming at the problem that the resolution of sperm in the sperm images of mobile phones and microscopes is too small,the data augmentation method is used to perform augmentation training on the target.And for the problem of insufficient semantic information of small target features,feature fusion is performed by multi-scale perceptual feature fusion.The multi-scale pooling module is used to establish the relationship between the front and the background,and the channel attention mechanism is used to suppress redundant features and improve the semantic information of the target.Finally,experiments are performed on three datasets to verify that the multi-scale fusion module can better integrate contextual relationships and improve detection accuracy.(2)Aiming at the problem that sperm morphology is simple and easy to overfit,an adaptive threshold regularization method is designed.The key information in the training process is randomly discarded for regularization,the attention mechanism is used to filter the target-related information in the feature map,and the adaptive threshold is used to screen the features with strong discriminative ability for random discarding regularization.The adaptive threshold can ensure that all the features of the target will not be discarded during each training,and only the features with strong correlation will be discarded to improve the expression ability of weak features,and solve the problem caused by the contradiction between the large number of parameters and the simple morphology of sperm.Fitting the problem so that the detector learns more robust features.Finally,it is proved by experiments that regularization can improve the detection effect to a certain extent,and it is proved by heat map that the boundary features are enhanced after regularization.(3)For the mobile terminal deployment problem of the detection model,the lightweight model is used as the benchmark network and the model is compressed by pruning,and the lightweight algorithm is deployed on the mobile terminal to realize a sperm detection mobile app based on deep learning.This paper proposes a multi-scale feature fusion detection algorithm for the sperm detection task,and finally implements a mobile sperm detection system based on Android,which can use the mobile terminal for sperm detection.Mobile sperm detection can effectively avoid the phenomenon of difficult sperm retrieval due to psychological reasons in patients in hospital settings.