Design And Implementation Of Object Detection Algorithm Based On YOLO

Object detection is one of the core topics in the field of computer vision and the basis of many other visual tasks.Its main task is to detect and locate specific targets from image information,and combines cutting-edge technologies in many fields such as image processing,pattern recognition,feature extraction,and deep learning,which is quite a challenging subj ect.As a common method of deep learning,convolutional neural networks(CNN)are widely used in the field of object detection because of their unique characteristics.At present,the object detection algorithm based on deep CNN has achieved a high level of detection effect,and has already achieved industrial application in some fields.However,due to the diversity of the environment in which the target obj ect is located,at present,there is no very good algorithm to achieve the effect of approximating human eyes.This thesis focuses on a CNN model based on YOLO algorithm,which has a certain improvement in both detection accuracy and training speed compared with the original algorithm.The main work is as follows:1.In the aspect of improving the detection accuracy of the algorithm,this thesis proposes two improved overlapping frame confidence strategy,linear attenuation function and Gaussian attenuation function.The experiment proves that it has a significant improvement on detection accuracy.This thesis also uses dynamic threshold design.Dynamically adjust threshold by using the density detection CNN.Setting a relatively smaller threshold at a denser picture position and setting a larger threshold at a lower density position can effectively improve the versatility of the neural network algorithm.2.In the aspect of improving the training speed of the algorithm,this thesis optimizes the loss function of the algorithm by analyzing the target loss function in the training process of the original algorithm.The loss operation of the wide and high parts of the bounding box modified by using the loss change rate to replace the original change value.This thesis also improves algorithm network structure by adding batch normalization layers at the input of the convolution layer to ensure that each layer of the network has the same distribution of inputs.At the same time,remove the dropout operation that was originally set to prevent overfitting during network training.This thesis uses 1*1 convolutional layer to replace the fully connected layer of the original parameter redundancy.3.Modularize the improved object detection algorithm for use in an actual obj ect detection system prototype.This thesis draws conclusions by comparing the experimental data of the object detection image database PASCAL VOC with the actual application performance of the algorithm.Due to the adoption of the above optimization designs,the improved algorithm has a more general application scenario and more efficient training efficiency.The improved algorithm based on YOLO has significantly improved detection accuracy in dense targets and small objects,and has wide application value.