The Research And Application Of FMCG Detection On Object Detection

Object detection is a very important and active research topic in the field of computer vision,and also an widely used technology,such as intelligent navigation,aerospace,manless driving,self-service store,etc.With the development of deep learning in recent years,object detection has been rapid developed both in theoretical research and real-world applica-tions.However the recognition rate of object detection is still influenced by small object,lots of objects in a picture,shield,the number of examples.Therefore,how to overcome the neg-ative effects of these factors,and to improve the accuracy and speed of the object detection are the huge challenges that the scientific researchers have faced.This paper proposes a densely connected residual network.The basic network of Faster R-CNN is used to extract the basic features of the input picture,and then used for the classification and locating of the object.Because the convolution network increasing with the number of network layers,the extracted features become more and more abstract,which is not conducive to the locating of the target.The residual network is usually used for the base network of Fast R-CNN.Each block of the residual network can extract a feature map,and the level of abstraction increases accordingly.In order to make use of each feature map of the residual network,each block of the residual network is connected to each other,and the connection mode is that,the input of each block is the output of all the blocks before it,and the output of each block is passed to the block behind it.Each block can directly obtain a gradient from the loss function and obtain the input from the first block.This connection method further reduces the problem of the gradient disappearing.At the same time,due to the full mix of features,the classifier can use all the features to classify,so that the classification effect is better.For the object detection problem,the feature contains the location information,it making the object's locating more accurate.In this paper,the network structure proposed in this paper is tested on the classification datasets CIFAR-10,CIFAR-100.The network structure has achieved very good results.At the same time,the imporved Faster R-CNN has been test on VOC2007 datasets,and the result is also exciting.This paper uses the Faster R-CNN framework to make a FMCG detection application.Currently,there are a large number of brands of beer in the market.Beer manufacturers need to analyze the sales of beer on the market every day to make decisions.The previous statis-tical method is that the sampler shoots the shelf picture and uploads it to review system,and then the reviewers review the beer picture.This manual method is not only error-prone but also very time consuming.The FMCG system automatically recognizes the beer in the pic-ture and gives the number of beer for each brand.The system adopts the algorithm of fuzzy judgment and tilt judgment to preprocess the picture,and adopts the method of multi-model fusion to improve the recognition accuracy.A single beer model and a packages of beer model were trained in the shelf scene,and a large-label model and a small-label model were trained in the ground scene.In the post process,misidentification of 330ML and 500ML beer types was automatically corrected.In the Faster R-CNN training phase,special settings were made for the anchor,and the anchor parameters were set according to the aspect ratio and size of the beer bottle.In the inference phase,tensorflow serving was used for model management,and Flask framework was used to provide WebService.The entire system can be deployed in distribution mode.