Stereo Target Recognition For Small Dataset Based On Generation Network And Classifiers Fusion

Artificial intelligence has entered a stage of rapid development in recent years.Target recognition is one of the most important research directions of artificial intelligence.The technology has changed from classifiers with image features to the deep learning.The performance of target recognition has been improved continuously in recent years.In addition to artificial intelligence,the sharing economy has fully integrated into our daily lives.The shared building blocks have gradually entered the field of vision.Due to the large variety of building blocks,the way of counting parts manually is inefficient.As a result,the technology using computer to identify building blocks has become a very popular research direction.The thesis focuses on a series of studies on the stereo target recognition and the problems encountered.In addition,the thesis takes the recognition of building blocks under arbitrary postures as examples.A brief overview is given as follows:(1)A stereo target recognition algorithm based on improved features and multi-classifiers fusion is proposed.The algorithm includes the improved histogram of oriented gradient and self-defined features.Based on the experiment of recognizing building blocks.It verifies that the algorithm is effective in the stereoscopic target with arbitrary attitude.In addition,the comparative experiments prove the superiority of the algorithm in this chapter.(2)Proposed a stereoscopic target recognition algorithm based on transfer learning and simplified network structure in the case of small dataset.The scale of training dataset has become an important aspect that restricts the performance of deep learning in practical applications.The thesis studies the problem of training networks in the case of small dataset.The algorithm uses the existing network model parameters as part of the initial value of the new network by transfer learning and builds a simplified network strcture as the rest of the new network.Finally,the experiments show that the algorithm can improve performance and can also achieve good results in the case of small datasets.What's more,the algorithm has a more streamlined network structure and smaller parameter scale.(3)The thesis proposes an algorithm using improved CycleGAN model to generate fake samples and the fake samples can assist small dataset training classification network.The algorithm merges the improved CycleGAN network and the convolutional classification network into a new network.The new network uses 3D model of stereo target and small dataset as input.What's more,it can work with end-to-end training.The experiments show that this algorithm can improve the accuracy of the classification network under small training datasets.The algorithm can generate high-quality samples too.In conclusion,the thesis mainly studies the stereoscopic target recognition algorithm based on improved features and multi-classifiers fusion.In addition,the problem of training network under the small dataset is also addressed.The high difficulty of training network is alleviated by transfer learning and simplifying the structure of network.Furthermore,an algorithm based on the fusion of improved CycleGAN and classification network is used to generate samples to assist training classification network.According to the study of recognizing building blocks,the algorithms proposed in this thesis has the advantages of good robustness and practical application value.There will be a great value for promotion in the future.