The Study Of Target Detection Algorithms Based On Deep Learning

Target detection mainly uses artificial intelligence technology to identify and locate objects in images.It has important applications in real life: intelligent monitoring of fire by digital camera,detection of tumors in medical images,automatic location of faces in digital cameras,etc.Traditional target detection can not make use of the deep features of the image,which often results in the detection results being easily disturbed by object occlusion,illumination changes,and so on.It is prone to miss detection and false detection.The emergence of deep learning has solved this problem well.Based on the study of target detection algorithms based on deep learning,this paper uses YOLO-v3 model to recognize various types of cells in blood cells and improves the accuracy of blood cell detection by improved algorithm.In order to improve the efficiency and accuracy of modern medical cell detection,the following two aspects were studied.(1)Firstly,the preprocessing technology of data expansion and image enhancement is proposed.This technique expands the training set by rotating and mirroring the image,and then enhances the details of the image by means of Laplace operator and histogram equalization.The results show that this method improves the detection accuracy,and the mean accuracy(mAP)of the three blood cells is increased from 0.5026 to 0.5722.(2)Secondly,the loss function of prior knowledge is applied to assign different loss weights to different types of samples.This can increase the penalty for the detected errors of low recognition rate categories.The optimal loss weight is determined by a series ofexperiments.The experimental results show that the improved loss function improves the detection accuracy of platelet cells on the premise of ensuring the detection accuracy of other two types of cells.The average correct rate(AP)of platelet cells increased from 0.2382 to0.5447.