A Study Of Motion Detection And Hand Writing Recognition

In recent years, with the improvement of computer operation speed, target detection and recognition by computer has been more and more widely used. Motion detection and handwritten digits recognition are two classic problems in detection and recognition. The traditional algorithms of motion detection perform poorly in the mixture of stationary and non-stationary scenes. There exists false alarm and ghost in the results. By reason of different feature spaces and types extracted by algorithms in handwritten digits recognition, the results rely on human’s experience and the accuracy is different.The principle and scope of moving object detection algorithm has been further studied. Aiming at the problem of high false alarm rate in background subtraction in the mixture of stationary and non-stationary scenes, an improved algorithm — BHZ-Mo G of Zivkovic’s Gaussian Mixture Model based on block histogram feature is proposed. First of all, an observation vector of image block is designed and blocks can be classified into static, dynamic and half-dynamic block by the statistical regularities of observation vector. Combining the observation vector and the classified information of a block, a method used to extract block histogram feature is presented. The block background model can be constructed and updated according to Zivkovic’s Gaussian Mixture Model and histogram feature. BHZ-Mo G can effectively reduce high false alarm rate of Zivkovic’s Gaussian Mixture Model for dynamic background. Experimental results show that BHZ-Mo G achieves higher precision than that of Zivkovic’s Gaussian Mixture Model while keeping the same recall. The maximal F1-scores are 0.7580 and 0.7902, respectively, for both Zivkovic’s Gaussian Mixture Model and BHZ-Mo G, showing that the presented algorithm can provide better subtraction results.The principle and calculation process of recognition algorithms in handwritten digits has been studied. To solve the problem of convolution neural network only extracting the common features in the same category of handwritten digits resulting accuracy of the special case is low, Sadowsky distribution and its probability explanation are presented. The searching algorithm of Sadowsky distribution and an initialization algorithm of convolution neural network based on Sadowsky distribution are designed.A learning model based on instance is proposed:firstly the convolution neural network is trained. Utilizing the searching algorithm of Sadowsky distribution and the initialization algorithm of convolution neural network based on Sadowsky distribution,the convolution neural network can be retrained. Combining the new convolution neural network and the original convolution neural network, the handwritten samples can be classified. Experimental results show that compared to traditional convolution neural network, convolution neural network based on instance has high accuracy for special cases.