Small Object Detection And Segmentation Algorithm Based On Convolutional Neural Network

The object detection and segmentation is one of the important research topics in the field of computer vision.Most of the traditional object detection and segmentation algorithms,which has a poor generalization ability,have to extract features in a specific task scenario by artificial design.Besides,due to the influence of fast illumination changes,different object shapes and occlusion between objects,the performance of traditional object detection and segmentation algorithms can no longer meet the needs of complex tasks.Because of the emergence of the convolutional neural networks,the shortcomings of traditional algorithms have been largely compensated,also with the increase in the accuracy and real-time performance.However,some characteristics of the small objects,such as less pixel information occupied in the image,unobvious features,and vulnerable to background interference,make the existing models less effective for the detection and segmentation and have obvious limitations.Therefore,it is a very difficult and meaningful research to deeply study the detection and segmentation algorithm of small objects.In this thesis,a new network structure and algorithm are proposed for the specific problems in small object detection and segmentation experiments:1.Small object detectionThis thesis deeply analyzes the advantages and disadvantages of the Single Shot Multibox Detector network,and explains the reasons for the poor detection of small objects.Aiming at the problem of missed detection and multi-overlap in small object detection,a network structure ELFSSD is proposed to strengthen the low-level feature information fusion.This structure is improved based on the SSD512 model.Firstly,the smaller feature conv4₃ is deconvolved to obtain the feature Dconv4₃,in order to increase the detail information of the small object in the feature conv4₃.Then,the lower layer conv3₃ and Dconv4₃ are fused by adding them point by point.Next,the fused feature is added to the prediction of the prior box to reduce the loss of the small object feature information.Finally,in order to solve the problem of imbalance between positive and negative samples,a focusing loss function is introduced.Based on the self-built rice detection data set,the average detection accuracy of the network structure is 91.2%,which is higher than the SSD model by 5.4%,and also higher than the DSSD model by 3%.2.Small objects Segmentation?1?In the process of small object segmentation with Adhesive rice as the research object,a concave segmentation algorithm of flat Angle difference is proposed for the over-segmentation problem caused by the pseudo-pits on the head of the rice.In this thesis,the adhesion region is selected according to the area and threshold of length-width ratio,and then the candidate concave points is found by the shortest Euclidean distance method.Finally,the candidate concave points are modified.According to the difference between the included vector angle and the flat angle of the concave point,the pseudo-pits of the rice head are removed,so as to improve the accuracy of the segmentation.This algorithm can achieve an average efficiency of 97% for the average segmentation of Adhesive rice.Although the algorithm greatly improves the segmentation accuracy,it requires high quality of the acquired image and is only suitable for small object segmentation with simple contour.For example,the nuclei in medical images are different in morphology and complex in contour.Therefore,this paper introduces the segmentation algorithm based on convolutional neural network to further study its segmentation on small objects.?2?The key techniques of the Mask RCNN segmentation algorithm for convolutional neural networks are studied in depth,and the instance segmentation of the nucleus is realized based on the nucleus dataset provided by the Mask RCNN in the kaggle2018 game.This thesis analyses the situation of missing segmentation and wrong segmentation in the color background image of data set,and puts forward the corresponding improved network.First of all,the atrous convolution is added to the network structure of the detection branch Res Net to prevent the loss of feature information while increasing the receptive field.Secondly,a right-side connection path is added to the FPN network structure to fuse more feature information,thereby improving the accuracy of feature learning.Finally,for the training method,part of the data is used to train the initialization parameters,and then the data set is trained as a whole.Compared with the original Mask RCNN network,the improved network structure raises the average accuracy of segmentation by 1%,and the segmentation speed is increased from 5fps to 7fps.