Image classification and detection is an important research topic in the field of computer vision.This paper first studies the image classification problem and then further studies the image target detection.In this paper,off-line handwritten Chinese character recognition is chosen as the background for image classification.For image target detection,this paper will analyze the existing target detection algorithm and its shortcomings,and improve it.This paper will study the offline handwritten Chinese character recognition and target detection based on the deep learning algorithm of convolutional neural network.The main research contents are as follows:(1)Due to the complexity and low accuracy of the traditional off-line handwritten Chinese character recognition process;However,the extraction of feature information in common convolutional neural networks is not sufficient.Therefore,this paper proposes a network model Character Net First of all,the feature grouping module is extracted through multi-level stacking to extract the deep abstract feature information of the image,and the communication and fusion of the feature information is carried out.Then,the downsampling and channel amplification modules are used to reduce the feature dimension while preserving the important information of the image.Finally,the feature information is refined and concentrated to solve the problem of overlapping and redundancy.In order to verify the validity of the model,CASIA-HWDB(V1.1)dataset containing 3755 Chinese characters was used to train and test the model.(2)Aiming at the low average detection accuracy of SSD target detection algorithm,especially for the detection of small targets,there are often problems of missed detection and false detection,in this paper,a multi-scale fusion target detection algorithm MFSSD is designed based on VGG network.In the algorithm,three kinds of convolution operations are used,namely,step convolution decomposition algorithm,deconvolution and expansion convolution down-sampling,to enhance the ability of extracting details and features of the network.Meanwhile,features of different layers of the network are fused to improve the semantic information of the upper and lower layers of the network.Finally,the VOC data set and the collected actual campus landscape images were used to detect SSD and MFSSD algorithms,respectively.The results show that MFSSD network effectively improves the comprehensive performance of SSD algorithm. |