LOGO recognition method based on attention mechanism image retrieval

Abstract

The invention discloses a LOGO recognition method based on attention mechanism image retrieval, and the method comprises the steps: judging whether there is a LOGO or not through obtaining an area containing the LOGO in an image, intercepting the area containing the LOGO in the image, and obtaining a feature tensor of the area containing the LOGO through a feature extraction network; and performing feature compression on the feature tensor containing the LOGO region to obtain a feature vector, performing feature space distance operation on the feature vector containing the LOGO region and thefeature vector of the LOGO in the image library, selecting an image with the shortest distance as the matched LOGO, reading a label corresponding to a feature code of the matched LOGO, storing the label in an image library, and determining the label of the LOGO by taking the label matched with the LOGO as the label of the LOGO to be identified. Under the logic, the situation that no LOGO exists can be filtered, the invalid operation process is removed, the operation efficiency is improved, and the problem that in the image retrieval process, the retrieval error rate is high due to the fact that the background proportion is large is solved.

Description

technical field [0001] The invention relates to the technical field of LOGO recognition, in particular to a LOGO recognition method based on attention mechanism image retrieval. Background technique [0002] LOGO (trademark / logo), which is a visual information expression formed by people in their long-term life and practice, has a certain meaning and can make people understand the visual graphics, and has a concise, clear and clear visual transmission effect . Traditional LOGO recognition methods include: method (1) realize a small number of specific LOGO recognition through target detection method alone; method (2) realize specific similar LOGO image search and matching through image retrieval method alone. [0003] Method (1) Target detection methods can be divided into two categories according to the processing flow: Two Stage and One Stage. TwoStage implements target detection in two steps. First, the sample candidate frame is generated through the feature network, and...

AI Technical Summary

Problems solved by technology

TwoStage implements target detection in two steps. First, the sample candidate frame is generated through the feature network, and then the candidate frame is classified through the classification network. Typical algorithms include Fast R-CNN and Faster R-CNN. The advantage is high precision, and the disadvantage is that the model is complex. Many parameters, slow inference

The One Stage target detection algorithm directly obtains the target category probability and bounding box through the output regression of the convolutional neural network. Typical algorithms include YOLO3, SSD, and RetinaNet.

The limitation is that the accurate classification of large-scale LOGO recognition cannot be achieved by using the target detection algorithm. Large-scale LOGO recognition involves thousands of types of LOGO, and target detection is to read the feature vectors of each dimension output by the convolutional layer. To obtain the category confidence information of the target, so the parameter amount of the target detection neural network is proportional to the number of target types that need to be recognized. Usually, the target detection model is only suitable for target recognition of dozens of types.

Therefore, if you use method (1) to do large-scale LOGO recognition, you will face two unsolvable problems: 1. The imbalance of category samples in the data set will lead to low classification accuracy of LOGO recognition, resulting in a large number of misjudgments

2. The parameters of the convolutional layer of the neural network are greatly increased, making the deep learning model very large, and the forward reasoning time is greatly increased, which makes the model unable to be deployed and applied

Its limitation is that the feature extraction is oriented to the whole picture, and the area inside the rectangular LOGO box in Figure 1 is effective feature information, and the rest is noise information. The noise feature information in the global feature vector obtained through feature engineering often accounts for Relatively large, in this case, feature encoding is performed, and the image retrieved after feature matching is often similar to the image background rather than the corresponding LOGO image

Images