Research On An Object Recognition Method Using Improved Convolutional Neural Network

In recent years,the convolution neural network has been successfully applied in the large-scale object recognition task with its strong feature generalization ability,and has caused the research upsurge of the related application.Large-scale samples and high-performance computers are two important aspects to improve the performance of object recognition.The existing improved convolutional neural network supports object regions of a sample to construct recognition model so as to improve recognition performance from the perspective of expanding the scale of samples.However,the performance of object recognition are still affected by the quality and quantity of object regions.Thus,in this paper,we improved the method of object regions extraction to study the object recognition of specific applications in unstructured environment,so as to construct a real-time and efficient object recognition method with strong expansion and low hardware cost.The work of this paper is as follows:The Selective Search algorithm extracts object regions containing redundant regions,so a new method named Hierarchical Selective Search is proposed in this paper to extract object regions,which is based on the region merging segmentation algorithm and improves the region merging criterion.In the process of region merging,the algorithm uses the combined similarity measurement composed from multiple complementary similarity measurements as the region merging criteria.As a result,it can reduce the redundant regions and get higher quality regions as objects or objects'parts search space.In order to verify the performance of the algorithm,a performance evaluation system is designed in this paper.The experimental results on two datasets showed that the proposed algorithm can effectively reduce the redundant object regions and ensure the quality of segmentation.While the time of object recognition is relatively long under the low hardware cost,this paper proposed a object recognition method for specific application using the improved convolutional neural network and the Hierarchical Selective Search.The method firstly use the Hierarchical Selective Search on every image to extract object regions for the purpose of reducing the cost in dealing redundant regions;then combine the methods of pre-training on large datasets and fine-tuning on a specific application dataset to train the parameters of the improved convolutional neural network to reduce the training cost and reuse models;finally use the fine-tuned network to extract the feature map of object region and construct the classifiers combined with regions tag information to recognize objects.The experimental results on the mobile product image dataset shows that the proposed method can effectively reduce the recognition time compared with the method based on the selective search;compared with the Bag-of-Word method,it has higher accuracy and scalability.