Research And Implementation Of Trademark Recognition Model And Trademark Instance Segmentation System Based On Deep Learning

Object recognition has always been a research hotspot in the field of computer vision,and trademark recognition is a specific application of object recognition research.In real scenes,trademark images are easily disturbed by natural conditions such as illumination and occlusion,which makes traditional image processing algorithms unable to meet the needs of accurate trademark recognition.At present,deep learning technology has achieved many breakthrough results in the field of target recognition,which can well solve the above problems.This paper uses deep learning technology to identify common trademarks in daily life,and combines image segmentation algorithms to build an instance segmentation system,which can be used for trademark retrieval or similarity detection.The main work of this paper is as follows:1.A multi-class trademark dataset for trademark identification is produced.The content of the existing public trademark datasets is outdated and cannot reasonably represent the common trademark categories in modern life scenarios,and the quality of images in the datasets is uneven,which will make the performance of the trained model poor.The dataset in this paper covers the common trademark categories in life,and also contains complex pictures with non-human interference factors such as strong light,overlap,blur,etc.,which improves the richness of the training data of the participating models and helps to improve the performance of the model in practical scenarios..2.A trademark recognition model based on hybrid feature pyramid structure is proposed.Feature images with rich information are the basis for accurate target recognition.The feature pyramid structure can be used to fuse feature information of different dimensions to obtain feature images with rich information,thereby improving the accuracy of target recognition.However,the traditional feature pyramid structure is top-down.The one-way fusion process cannot fully integrate multi-scale information,which will eventually affect the accuracy of target recognition.To solve this problem,this paper proposes a hybrid feature pyramid structure,which cross-fuses multi-layer feature images,and performs bottom-up fusion after cross-fusion,so as to obtain feature images with rich information.The trademark recognition model using this structure is tested on the dataset of this paper.The results show that the m AP of the model is 97.8%,which is 6% higher than that of the trademark recognition model using the original feature pyramid structure.In addition,experiments are conducted on the public dataset Flickr Logos-32,and the m AP of the model in this paper is 87.4%,which is better than the Faster RCNN,PANet and DSFPN models.3.Combining the trademark recognition model with the improved Grab Cut segmentation algorithm,an interactive trademark instance segmentation system is constructed.In addition to the accuracy of segmentation,the trademark instance segmentation system also has certain requirements for real-time performance.In order to improve the real-time performance of the system,a lightweight Mobile Net V2 convolutional network is used to improve computing efficiency,and Soft-NMS is used instead of NMS for better screening.For the candidate box,the cosine annealing algorithm is introduced to adjust the learning rate and optimize the gradient descent process.The instance segmentation system constructed in this paper is tested on the dataset of this paper.The results show that the segmentation accuracy of the segmentation system is 92%,and the processing time of a single image is 0.39 s,which is shorter than the Mask RCNN model.