Research On The Application Of Image Recognition Technology Based On Deep Learning Algorithm

Image recognition, as an important branch in the field of pattern recognition, has been widely used in the field of computer vision. Since 2012, the image recognition method based on deep learning, especially convolution neural network, has made a number of major breakthroughs.Various types of models have been kept springing up. Technological applications of image recognition are also overwhelmed, such as face recognition, fingerprint recognition, traffic violation camera and so on.Recognition accuracy and detection speed are of high importance to Image recognition, which directly affect image recognition technology on practicality and feasibility. Compared with other traditional recognition methods, deep learning algorithm can extract deeper image features. Under the support of massive data, deep learning algorithm can express the model more effectively. In this paper, the author combines the deep learning algorithm and image recognition technology to build a model for image recognition to further improve the performance of image recognition and image detection meanwhile improve the training process.The main work of this paper is as follows: First, image recognition technology based on deep learning algorithm is applied to the detection of cancer cell, which is used for pathological analysis. By improving the AlexNet network model and adding the feature normalization layer, a series of data preprocessing and model training methods are adopted to make the network training more convenient and stable, and the recognition accuracy of the algorithm is improved effectively. Secondly, based on the analysis of cell image, the convolution neural network model is compressed. By pruning and retraining the model, the model is reduced to the original 13%,and the detection speed is increased to 3 times, which can improve the practicability of the model. In addition, this paper compares the influence of different pruning proportion on the final result of the model, and finds the optimal pruning proportion and compression ratio.Thirdly, an improved object detection model SSD is proposed. In this paper, cascaded deconvolution layers are added to the SSD model, and five new feature layers are constructed, which increases the feature richness and the receptive field size of each candidate box, effectively improves the object detection accuracy.The main innovations and contributions of this paper are as follows:In this paper, an improved AlexNet compression network is proposed and applied to the task of cancer cell detection, which improves the detection accuracy, reduces the model size and improves the detection speed and efficiency. Finally, cascaded deconvolution layers are added to the object detection model SSD, and five new feature layers are constructed, which effectively improves the accuracy of object detection.