Research On Deep Learning Based Image Recognition Application

Image recognition is an important direction in the field of artificial intelligence. After years of research, image recognition technology has made some progress. Image recognition technology mainly consists of feature extraction and classification, and feature extraction is the bottleneck problem of image recognition technology, which directly determines the recognition performance. Image features are divided into global features and local features, but these features are underlying visual feature of the image, and requires a certain expertise personnel to design and chose. This artificial designed features need to go through a lot of validation to prove its effectiveness on a particular recognition task. To a certain extent, it has limit the application of image recognition technology.In recent years, with the advent of the era of big data and computing resources are getting cheaper, deep learning technology has been continuously developed. It is a datadriven model, learning feature from image recognition tasks with large amounts of data. To this end, in this paper the theories of deep learning are firstly studied, then we expand these theories to the pornographic image recognition and license plate recognition. The main contents of this paper includes the following sections:Firstly, we study the deep learning technology. Focus on the feature extraction of deep learning and three important network models. Deep belief networks(DBN) can achieve unsupervised learning feature; Convolutional neural network(CNN) has been widely used for image recognition task; Recurrent neural network(RNN) can learn feature from the sequence data. We further study these network structure and training methods.Secondly, this paper presents pornographic image recognition method based on CNN. The method consists of two steps, coarse and fine detection. Coarse detection can quickly identify image with no or less skin color and mug shot, because most of the images are normal images, which can greatly reduce the recognition time. For more area containing skin color images, further fine detection is need. Firstly we use a large number of labeled pornographic and normal images to train CNN classification model, and then use this model to recognize pornographic images. We conduct the experiment with a database containing 19000 images, experimental results show that recognition accuracy of the proposed method can reach 97.2%, much higher than traditional pornographic image recognition methods.Finally, the traditional license plate recognition algorithms contain processes of license plate location, license plate correction, character segmentation, and character recognition, each process will affect the license plate recognition accuracy. This paper presents a license plate recognition algorithms based on convolutional recurrent neural network(CRNN). This method consists of plate location and plate recognition. Firstly, license plate location based on the edge is used to get license plate candidate region, and then the CRNN is trained. Last, employing the CRNN model to recognize plate. This algorithm is an end-to-end method without license plate correction and character segmentation, which provides another idea for license plate recognition. Experimental results show that, recognition accuracy of the proposed method can reach 76%. The plate without provinces character recognition accuracy was 91%.