Research On Image Object Detection Algorithm Based On Convolutional Networks

Object detection algorithm is an important branch of computer vision.The algorithm based on convolutional neural network has been widely used in many fields such as vehicle detection and face recognition.However,it was not applied well in some fields such as medical image detection.Object detection algorithm can be efficient in assisting doctors to grasp patient's condition,thus promoting the development of intelligent medicine.Therefore,research on algorithms for improving the accuracy of small-scale object detection is necessary.There are many kinds of object detection algorithms at present,and the two-stage object detection algorithm,which contains region proposal stage and feature extraction stage,is the most accurate one.Region proposal stage is used to generate a large number of region proposals to search locations which contains object instances.Feature extraction stage uses convolutional neural networks to extract features and then classify the region proposals.The two stages are combined to achieve the object location and recognition.There are two problems with the two-stage object detection algorithm.First of all,it is necessary to collect a large amount of annotated images to train the convolutional neural network,otherwise it will lead to over-fitting.However,it is often impossible to ensure sufficient data in practice.In addition,recent algorithms have insufficient use of convolution features,making it difficult to achieve satisfactory results in the small-scale object detection such as medical images.In view of the above problems,this thesis puts forward a new object detection method based on data augmentation and multi-scale feature integration.The main research results are as follows:1.Aiming at the problems of low quality,distortion,and inconsistency with the original image style for images generated by traditional data augmentation algorithms,this thesis proposes a new data augmentation algorithm based on image blend and style transfer.The algorithm first combines various image blend methods to generate high-quality blend images based on a small datasets.And then uses the modified Cycle-Consistent Adversarial Networks for style transfer,generates new data which is consistent with the original image style from the blend image.Through the combination of these two methods,the data augmentation algorithm can generate a large number of simulation data suitable for object detection,thus solving the problem of over-fitting when using small datasets for training.2.This thesis proposes a new object detection algorithm based on multi-scale feature integration.The algorithm has two improvements: First,this thesis proposes a new feature pyramid structure.A bottom-up pathway of feature integration is used to enhance the expression of low-level convolution feature information,which can improve the ability of small-scale object detection.At the same time,multi-scale features of the feature pyramid structure will be used simultaneously to ensure the expression of semantic information from high-level convolutional feature.Secondly,the algorithm uses the distribution information of bounding boxes as weights to fix the location of the bounding boxes in the region proposal stage to make the location of region proposals more accurate thus improving the location ability of the algorithm.The precision,recall and average precision of the modified object detection algorithm trained by augmented medical image dataset reach 77.3%,88.3% and 65.1% respectively,which are better than the other object detection algorithms.3.Based on the modified algorithm,this thesis designs and implements a object detection system,carries out the system design and implementation,and shows the test results of the system in detail.