Research On Neural Network Architecture Design For Image Segmentation And Classification

The design of neural network architecture in deep learning is crucial for optimizing model performance and effectiveness.Neural network architecture design involves selecting appropriate network structures,layers,neuron quantities,and designing connection patterns and activation functions to effectively address specific problems.Varying problems and tasks may necessitate distinct neural network architectures.Image segmentation and classification tasks often utilize deep convolutional neural networks(CNN).The architectural design of neural networks plays a vital role in deep learning,as making appropriate selections and designs can effectively extract feature from input data,optimize model parameters,and improve the model’s generalization ability and computational efficiency.The main objective of this research is to address image segmentation and classification tasks.It explores various neural network architectural design approaches,including object-scale-oriented and multipath-based convolutional neural network models,hybrid neural network models that integrate position priors and level set segmentation,and growable architecture convolutional neural network models based on incremental branch growth methods.The primary research achievements of this dissertation can be summarized as follows:1.To tackle the issue of object scale sensitivity in image segmentation using convolutional neural networks,we propose a fully convolutional network with an object-scale-oriented design that incorporates multiple paths.This novel network architecture is specifically tailored for the segmentation of pigmentary skin lesion images.Given that the scale of pigmented skin lesions in the training dataset can impact the segmentation results during testing,a training strategy named object scale-oriented training is proposed.First,the pretrained VGG-16 network is adapted and converted into a fully convolutional network.Subsequently,the fully convolutional network is fed with pigmented skin lesion images containing boundary annotations after undergoing simple preprocessing.Fine-tuning training is then performed using a pretrained model trained with the object scale-oriented training.During the object scale-oriented training process,the training dataset is partitioned into two subsets according to a metric known as the object occupation ratio.Subsequently,three fully convolutional networks,each oriented towards a different scale,are trained using the complete training dataset and the two subsets.Training and testing rely on the ISIC 2016 dataset,made available by the International Skin Imaging Collaboration(ISIC).Comparative experiments with state-of-the-art algorithms validate the proposed algorithm’s superior or equivalent segmentation accuracy.This method exhibits the following characteristics: firstly,a coarse classification network trained on the entire training dataset is employed to estimate the scale of objects,thereby enhancing object scale estimation accuracy during testing.Secondly,a fine-grained classification network trained on samples from a specific scale range is chosen for final segmentation based on the estimated object scale,thereby improving object segmentation accuracy.Moreover,this method solely necessitates simple preprocessing of sample images and does not incorporate any postprocessing.2.In order to mitigate the significant loss of positional information in encoding-decoding convolutional neural networks,we introduce a hybrid architecture deep convolutional neural network that automatically incorporates positional priors.Our proposed approach is successfully applied to the segmentation of pigmentary skin lesion images.The model combines CNN with a level set segmentation model for image segmentation.First,we proposed a backchannel filling method.This method fills the estimated target position by the initial forward propagation of the CNN into the prior position channel,named as idle channel.Next,we established an end-to-end deep learning model that combines a multi-branch encoding-decoding CNN with level set.This model predicts the initial level set and drives energy through the second forward propagation of the network to complete the evolution in the level set segmentation model.Additionally,we designed a learning algorithm that combines a multi-branch encoding-decoding CNN and level set to train the segmentation model.Unlike the traditional interactive labeling input prior,the multi-branch encodingdecoding CNN takes the image plus an idle channel as input.It estimates the target position during the first forward propagation of the CNN,and then fills it back into the idle channel to automatically acquire the prior position.Second,the second forward propagation outputs the energy that drives the evolution of the level set and the initial position.We compared the experimental results of the proposed algorithm with those of other state-of-the-art algorithms using the ISIC 2017 and ISIC 2016 pigmentary skin lesion image datasets.The results demonstrated that our method achieved higher accuracy compared to other algorithms,with Jaccard indices of approximately 77.0% and 85.1%,respectively.3.To address the limitations of existing neural network architecture search methods in terms of requiring complete retraining and high computational complexity for image classification,we present a novel approach based on the incremental branch growth convolutional neural networks.By adopting this method,we effectively reduce the computational burden while achieving accurate surface defect identification in piezoresistor image analysis.IBG-CNN is an automated approach used to construct CNN.Initially,this method establishes an initial network based on incremental branching and proceeds with training it.Subsequently,the network continues to grow through an iterative process known as "branch insertion-incremental training." Branch insertion involves identifying incremental branches that can be incorporated into the existing network.Subsequently,a new branch is constructed to connect to the incremental branch in a bottom-up manner.Incremental training involves fine-tuning the weights of the previous generation network and re-training the new branches to accelerate the network model search process during the growth process of the next-generation networks.Training and testing were conducted on six classifications representing three types of defects in piezoresistors,based on IBG-CNN.The automatic construction process of the entire network was performed using only one GPU and lasted approximately 16 hours.Finally,we evaluated the test dataset consisting of 6248 images using the mean Average Precision(m AP).Our algorithm achieved a higher detection accuracy(m AP ≈ 0.929)compared to two other state-of-the-art CNN methods,as indicated by the experimental results.