A research team from the Engineering Research Centre of Applied Technology on Machine Translation and Artificial Intelligence, Ministry of Education at the Macao Polytechnic Institute (MPI) has investigated the concatenate-designed neural network model and its method of feature extraction from images, designed classifiers to collect features produced between network sets, and obtained results from the classifiers through the proposed normalization method for better classification. The research, contributed by PhD student Ka-Hou Chan, Professor Sio-Kei Im and Associate Professor Wei Ke, can be effectively applied to the tasks of recognition and classification and has significantly improved the accuracy and stability of the application. The article has been published in the internationally renowned artificial intelligence journal Neural Computing and Applications.
Most applications of intelligent networks for image analysis are designed with a large network architecture by stacking numerous layers. Since convolutional layers are effective in feature extraction from images, most applications are based on stacking a large number of convolutional layers to extract the major features from original image information. However, convolution layers cannot take into account the structure of the image, and the entire information of target would be diluted or discarded during several convolutions, thus reducing the accuracy. The traditional design is to connect the information before convolutions to the final classifier as references for decision making, but this increases the workload of the final classifier and dilutes its neuron performance. In particular, when receiving high-resolution images, more convolutional layers are required and the structure of image will be largely lost, causing the process and accuracy of machine learning to become unstable. MPI’s research team introduced multiple classifiers to achieve feature analysis of each convoluted layer, with the earlier set obtaining the entire information, the later sets identifying the local features of images, and each classifier summarizes the current features as a reference for a final decision. Therefore, the proposed method can alleviate the pressure of the final classification. After development and experimentation, the team introduced a novel normalization method that aims to combine the result of each classifier. The proposed normalization is effective in reducing the training time of each model in view of the large amount of machine learning processing time required to use multiple classifiers, and also demonstrates the performance of convergence enhancement.
During the research experiments, significant improvements were found in the learning results of using multiple classifiers, especially for the dataset with a large number of categories, and the proposed normalization method was found to provide more accurateresultsfor image classification. The research work was funded by The Science and Technology Development Fund, Macao SAR (File no. 0001/2018/AFJ). Full text available at: https://link.springer.com/article/10.1007%2Fs00521-021-06462-0View gallery