Research And Implementation Of Image Recommendation Algorithms Based On Deep Learning

With the rapid development of information technology and business platforms,the amount of image information in e-commerce platforms is growing rapidly,and the demand for image content information in scientific research and business applications is becoming more and more intense.For e-commerce platform,it can effectively improve the service quality and competitiveness of e-commerce platform by quickly finding out the products images that users may be interested in from the massive image database and recommending them to users.The traditional image recommendation algorithm uses text-based method,which is greatly influenced by human subjective factors and photographic environment.How to improve the recommendation efficiency of e-commerce image has become an urgent problem to be solved.In recent years,deep learning has become a hot research direction in the field of machine learning,and has been widely used in image classification,target detection and natural language processing.In this paper,image recommendation algorithm based on deep learning is studied,and an image content-based recommendation algorithm is designed.The algorithm mainly includes three deep neural networks.The goal is to use deep learning technology to make up for the shortcomings of traditional text-based image recommendation methods and to reduce the impact of human subjective factors and shooting environment.The main contents of this paper are multi-attribute image classification,optimization of target detection algorithm and dimensionality reduction algorithm based on convolutional neural network features.On this basis,a prototype of image recommendation system is designed and applied to e-commerce image recommendation.The main innovative achievements of this paper are as follows:(1)A multi-task learning method based on improved convolutional neural network is studied and designed.This method draws on the idea of parameter migration in transfer learning,and is used to solve the problem that traditional convolutional neural network cannot classify multiple attributes contained in commodity images at the same time.Aiming at the problem that there are fewer specific types of goods in the data set and there are imbalances in categories,an over-sampling strategy based on Mixup algorithm is proposed.At the same time,the relationship between the complexity of image attributes and the sparse rate of CNN output characteristic matrix is studied,and the improved Grad-CAM algorithm is used to visualize and analyze the key areas of image attributes recognition,which improves the interpretability of the network.(2)A Faster-RCNN algorithm based on adaptive pooling for commodity target detection is studied and designed.The goal is to solve the problem that traditional pooling method cannot effectively extract the details of wrinkles and textures in e-commerce images.The pooling algorithm is improved based on the traditional maximum pooling model.Experiments show that this algorithm has better recognition effect for commodity images with complex texture and wrinkle information.(3)The multilinear principal component analysis method is applied to dimensionality reduction of feature extraction from convolutional neural networks for the first time in this paper.The aim is to solve the problem that the dimensionality of image feature extracted by convolutional neural networks is too high and they have correlation.On this basis,an image hash coding method based on deep learning is designed for image retrieval combined with local sensitive hash algorithm.Experiments show that compared with the traditional dimension reduction method,the multi-linear principal component analysis method has better dimension reduction effect for extracting features from convolutional neural networks.(4)A prototype system of image recommendation is studied and designed.The system is based on the three deep neural networks designed in this paper and is written in Python language.In this paper,the system is applied to image recommendation of e-commerce,and good recommendation results are achieved.