Research And Application Of The Convolution Neural Network

Convolution neural network (CNN) is a multi-layer perceptron that specially designed toidentify two-dimensional image, it has good fault tolerance, parallel processing capabilities andself-learning ability. In dealing with the problem of two-dimensional image, especially in theidentification of displacement, scaling and other forms of distortion invariance, it has goodrobustness and computational efficiency. It has been widely used in face recognition, eyedetection, pedestrian detection and robot navigation.In this paper we focused on the theory of CNN which based on neural network. And basedon this theory algorithm we took further development of the applications in handwritten numbersrecognition and wood defect recognition. The main work was described as follows:1.Analysed the network structure of the CNN with regard to the shortcomings of CNN, weproposed a algorithm which can constantly adjust the training set for dealing with themisclassified samples;2.A progressive expansion of network structure algorithm was given which can adjust thenetwork structure automatically let the network structure adapted to the practical problems andreduce unnecessary network layer;3.As the CNN on the feature extraction of the sample is adaptive it can make the operationsimple however it will bring numbers of issues like feature dimension reduction and otherpretreatment a improved CNN was proposed which made the feature extracted by the networkricher, and made the feature extraction in Non-network self-extraction, also can analysis of thedata in the network specifically;4.Made CNN application of handwritten numbers recognition in Gray-scale image andwood defect recognition in Color image. Gave CNN algorithm in actual recognition applications.Contrast CNN and mainstream classification algorithms like Clustering Support VectorMachine Adaboost Neural Network and so on. Experiments showed that the accuracy andefficiency are superior to the other several algorithms.