Design And Implementation Of Taobao Image Quality Classification System Based On Deep Transfer Learning

Image processing is one of the tasks that Taobao shopkeepers must do before they put their products on the store.Images with visual impact will immediately arouse consumers' desire for consumption.At present,many shopkeepers,especially new shopkeepers,often encounter not only the problems that the image is too large to upload,but also the problems of unsatisfactory sales due to the low-quality images caused by inappropriate brightness,definition and other factors.Some of them even have no way to deal with the image.At present,there are few image processing tools that can match Taobao's image standard to assess by one click in the market.Image correction is often completed by complex parameter adjustment.In order to solve the above requirements,this paper will use deep transfer learning technology and blind image quality assessment to achieve the automatic batch assessment of images.At first,the data set of Taobao image is acquired by the web crawler,then the image is processed and Res Net50 is fine-tuned.After training,a new neural network is acquired for the image classification system.In order to solve the problem that some images are assessed as the qualified by the system but the actual effect is not good enough in the test,an improved method is proposed,which focuses on the improvement of the method to find out the threshold of brightness and definition,so as to realize the optimization of the system.The research work will be as follows:(1)Improve the super-pixel algorithm based on the brightness.Based on HSV color space,we build a forest first,then arrange the connected edges of all pixels in ascending order according to the weight(brightness value difference),then calculate the internal difference of the corresponding area,and then traverse all the edges in the graph.If the brightness weights of two adjacent regions are less than the internal difference of their respective regions,then the two regions are merged.(2)Establish the data set.After building a Taobao image data set by online crawler,the improved super-pixel algorithm and the variance formula of Laplacian transform areused to find out the better threshold of brightness and definition by experiments.The data set is divided into two parts with the help of the threshold,which are the qualified and unqualified.(3)Train a neural network model.Based on the improved data set,fine-tune Res Net50 neural network model and train a new neural network model which is suitable for Taobao image assessment.(4)Develop an image quality classification system.With the help of the pre-trained neural network model,an image quality classification system is designed and programmed.Its main function is realized by a module for batch comprehensive assessment,its auxiliary functions are realized by a module for better Taobao images,a module for size change and word adding,a module for compression and a module for background service.It is tested in the real environment,and the experimental results are in line with expectations.