Multi-task Deep Learning Product Image Classification Method Based On Convolutional Neural Network

As the basis of artificial intelligence,deep learning has a wide range of application fields,especially the convolutional neural network has achieved significant research results in computer vision and other fields.In order to adapt to real life scenes,the form of deep learning tasks has gradually changed from single-task to multi-task.An effective multi-task deep learning model can use information commonality and differences among multiple tasks to improve the learning performance of one or several tasks,in-depth research on multi-task deep learning will help promote the development of various fields.With the growing e-commerce market,the products become more and more diverse,meanwhile,it brings great challenges to the rapid and accurate information labeling of products by platforms and merchants,the multi-task product image classification technology combined with attribute prediction and category classification has emerged to solve the problem.An effective multi-task product image classification model can provide technical support for services such as product image category labeling,product retrieval and product intelligent recommendation.This thesis aims to build a multi-task deep learning model for product image classification,mainly focusing on two aspects:convolutional network feature transfer strategy formulation and multi-task parameter sharing mechanism improvement:(1)Most of the convolutional network transfer learning research only uses the pre-trained model as a feature extractor,and there is a lack of meticulous research on the pre-trained model.To address this deficiency,this thesis builds a basic convolutional network feature transfer model for multi-task product image classification,and the pre-bottleneck layer network and parameters of VGG16,InceptionV3 and ResNet50 are transferred into the basic model to examine the generalization ability of the three pre-trained models on the product image data in this thesis;Then this thesis selects the pre-trained model with strongest generalization ability,divides it modularly,and uses the connection point between each module as the segmentation node of transfer feature freezing and retraining in turn to study the impact on model performance when the transfer feature freezing and retraining segmentation nodes are different.In this thesis,DCSA and DeepFashion product image datasets are used for experiments.The experimental results show that the VGG16 pre-trained model has better generalization ability on these datasets.In addition,for different datasets,there may be different transfer feature freezing and retraining optimal segmentation node.This part of the research results can provide some references for the formulation of transfer strategies in convolutional network transfer learning.(2)To address the problem of insufficient retention of task-specific information before branch node and insufficient interaction of task-related information after branch node in the commonly used soft and hard parameter sharing mechanism for multi-task deep learning,a partially shared unit which can be embedded in multiple isomorphism networks has been designed in this thesis to maximize the retention of task-specific information as well as task-related information.This thesis finally combines the partially shared unit and convolutional network feature transfer strategy to build a multi-task deep learning model and apply it to the product image classification.The model is verified using DCSA and DeepFashion product image datasets,and then this thesis compares the learning results of this model with the learning results of isomorphic hard parameter sharing multi-task network and soft parameter sharing multi-task network.Finally,it is proved by experiments that the model constructed in this thesis has a better classification accuracy,indicating that the model in this thesis can mine the information between tasks more thoroughly than the traditional soft and hard parameter sharing mechanism.