Study on the Optimization of CNN Based on Image Identification
Yanyan Feng, Shangyou Zeng, Yang Yuan-fei, Yue Fei Zhou, Bing Pan · 2018
The feature extraction method of traditional image classification is difficult to deal with complex image problems. Although feature detector based on convolutional neural network can easily extract image features, many current network models have poor recognition accuracy and too many parameters. This paper proposes a multi-scale dual-channel dimension reduction module (DR module) to extract image features. Based on the AlexNet model, a deep global optimization model (GONET model) is proposed by exploiting the DR module, dropout and global pooling strategy. Compared with the AlexNet model, this model has better recognition performance. The accuracy on the Caltech256 dataset reaches 58.8% with GONET model, exceeding that of the AlexNet model by about 4.0%. The accuracy on the 101_food dataset reaches 69.0% with GONET model, exceeding that of the AlexNet model by about 8.9%. The experimental results show that the GONET model has superior image recognition effect, and it significantly reduces the network parameters while improving the recognition accuracy.