Multi-scale orderless pooling of deep convolutional activation features

I. main ideas in this article

Due to the lack of geometric immutability of global CNN features, the classification and matching of variable scenes are restricted. Therefore, this paper aims at this problem and does not reduce the CNN capability, multi-scale orderless pooling (MOP-CNN) is proposed, that is, the unordered CNN of multi-scale image extraction (three levels) is obtained. Then, after PCA dimensionality reduction, vlda encoding, PCA dimensionality reduction, finally, the features of these three levels are cascaded into 3*4096 dimensions, and linear SVM classification of One-vs-all is adopted, which achieves certain results. The Feature Extraction structure is shown below:

2. Multi-scale orderless pooling (MOP-CNN) Framework

This article uses the open-source caffe CPU mode to run

1. Scale the original image to 256*256, subtract the mean value, and then input caffe to extract the 4096 dimensions of Layer 7 as the level1 feature;

2. In the original image, the window is 128*128, the step is 32, and m images are added to the batches. Each batch is extracted based on the method described above, assume that M features of 4096 dimensions are generated; PCA is used to reduce 4096 dimensions to 500 dimensions; k-means is used to train a 100*500 codebook; we use the simple version of vglad to encode M 500-dimension features and normalize the two-norm to generate a 50000-dimension feature. We use PCA to reduce the dimension to 4096, the level2 features of the last 4096-dimension feature are left and right;

3. The method is the same as the preceding step, except that the window size is 64*64, and the level3 features are left and right about 4096 of the features in this step;

4. normalize the 4096 features of these three levels into a 3*4096 dimension vector feature;

The vlda encoding method is as follows:

5. One-vs-all linear SVM is used for classification.

Iii. Summary

The idea of this article is simple and clear, and the multi-scale features are combined with CNN, using the vglad encoding method. In terms of this method itself, the multi-scale method has achieved good results compared with the single-scale method. This article is based on multi-scale feature extraction at the image layer. The next article will briefly introduce a pooling paper on the feature layer.