Rearch On Image Edge Detection And Object Classification Based On Cross-Fusion Neural Networks

In recent years,with the rapid development of image processing technology,image edge detection and target classification occupy an important position in the field of computer vision,and are also research hot spots in the field of image processing.Accurate image edge detection and target classification have a great influence on high-level feature extraction,feature description,target recognition and image understanding.Image edge detection and target classification technologies are mainly divided into traditional detection techniques and methods based on gray level extraction,and detection techniques based on convolutional neural networks.The use of a Caffe framework and a VGG16 template to propose a new accurate edge detector based on the convolutional neural network in the detection and extraction of image feature points was studied in this thesis.The research on object multi-classification algorithms is improved based on Faster R-CNN.The image target classification and recognition model not only accelerate the efficiency of image edge detection,but also improve the accuracy of image edge detection and the accuracy of multi-class positioning of objects.The main research work of this thesis includes:1.Aiming at the shortcomings of traditional feature point extraction algorithms such as low edge extraction accuracy and susceptibility to noise interference,the current domestic and foreign latest convolution feature point extraction algorithms are searched.Finally,a five-layer classic structure network of deep learning convolutional neural network VGG16 was used in thesis.The maximum pooling layer was used between each combined neural network layer to restore the scaled feature image as much as possible.When extracting the feature image,the VCF neural network structure was proposed in this thesis,it was used to extract the feature points at the edge of the image.2.The VGG16,HED,RCF and other renderings were obtained from the edge detection of the image extracted by the convolutional neural network.They were too dependent on the fully connected layer to have a certain degree of rough edge lines.And because of the loss function was set improperly,it was easy to cause problems such as disappearance of the gradient and loss of a large amount of main feature information.In this thesis,the gradient dimensionality reduction method of multiple convolution kernels is used to complete the acquisition of low-level and high-level feature objects in two directions.Then through the top-down method and the circular convolution flow from left to right,it can ensure that the loss function of each layer can be relatively stable forward learning and backward feedback.3.After completing the convolutional feature multiple times,the target image and the reference image are mapped to the same coordinate system and processed by a set loss function and fusion layer.The third layer of lateral output was used as the base to perform cross-network fusion with output maps of layers 1,2,and 4 respectively.And the loss value of each layer was calculated by the weighted loss function of the fused feature image.If there was a loss of the learning target or the original position gradient cannot be traced back in the deconvolution process,the training was stopped and the disappearing layer was reset until the end of the loop training,and the model gradually matures.It can make the final result closer to the target edge curve.4.When multiple targets were on the same image,the network can still capture the features of the target and locate it,so that the confidence score can reach above 0.8,and accurate target classification can be performed.In the RPN cross-layer output result,the previous feature information between 0.5 and 0.7 was fully used.The result contains a rich of positive sample objects,and a lot of non-interesting information is eliminated,so the object classification of image was getting clearer and more accurate.