Design And Development Of Image Recognition System For Fast Selling Products Based On Deep Learning

With the economic and social development,the fast-selling goods industry has developed rapidly,and the beer industry has become increasingly fierce.How to seize the market to obtain first-hand market share information has become a problem for beer manufacturers represented by Budweiser Beer.At the same time,in the era of Internet big data,deep learning has developed rapidly and it has begun to be widely applied in all walks of life.How to use deep learning techniques to quickly count the number of different types of scenes selling beer and competitor beer sales in major supermarkets in order to provide accurate evidence and provide a basis for priority consideration.This paper cooperated with the Budweiser beer manufacturer and used the Faster RCNN object detection technology to identify a total of three types of beer(shelves,freezers,and ground piles),different types of self-selling beer,and competitor beer sales for a total of 76 types of beer.Providing image pre-processing(image correction technology)for images with more slanted images;adding post-processing to the model for repeated frames;Adding multi-model fusion for oblique and blurred images,which greatly improve image recognition accuracy rate.Providing a set of solutions for Budweiser brewers,including image recognition,page display,and interface development.It greatly shortens the time for manually processing pictures,improves the accuracy of image recognition,and provides powerful support for the promotion of the later stage.The main work is as follows:(1)A variety of target detection algorithms were analyzed and compared.Faster RCNN with faster speed was used as the target detection algorithm.The label image software source code was modified to improve the efficiency of plotting.Based on the VOC2007 dataset,multiple networks were used for training,such as VGG16,ResNet50,Inception-v2 and so on.(2)Providing image processing tools written in C++ for images in which image tilt is severe.First,looking for two up and down shelf lines and determine the two vertical lines depending on the orientation of the beer bottle,which will be in the inclined trapezoidal image part and mapped to the two-dimensional coordinate pixels through the coordinates.For the image outside the trapezoid,the original pixels are kept,so that image correction is realized.(3)It mainly solves the three problems of inaccurate identification:a.Repeated frame problem,b.Types with more errors,c.Different heights and sizes of the same type name.Using multiple sets of samples to train to obtain multiple models,such as the positive and negative sides were obtained models.For the same category,the two positive and negative recognition results are unified into one category.The image merging operation is performed for IOU between more than 50%between different positions of the frame.For the same height of the different type name,such as 330ml and 500ml,taking the average of the same row of beer bottle heights,beer bottles that exceed the average of 20%are considered to be 500ml,and vice versa.(4)The client or web page uploads a standard image to Microsoft Azure(Microsoft Cloud Platform).The server downloads the standard image,periodically uploads the result of the review,and the client or web server obtains the result of the review.The WebService framework is mainly used.The model training end is published as-a service using Python.The client or web page requests service via Ajax,and the recognition result is obtained and uploaded to the Azure cloud.