Multi-layer feature fusion fine-grained image classification method with parallel convolution blocks
A technology of feature fusion and classification methods, applied in neural learning methods, instruments, biological neural network models, etc., can solve problems such as high computer requirements, small number of samples, complex network models, etc., to enhance interaction capabilities and enhance feature expression capabilities. Effect
- Summary
- Abstract
- Description
- Claims
- Application Information
AI Technical Summary
Problems solved by technology
Method used
Image
Examples
Embodiment 1
[0041] A fine-grained image classification method based on multi-layer feature fusion with parallel convolutional blocks, see figure 1 , the method includes the following steps:
[0042] 101: Obtain internationally public fine-grained datasets;
[0043] 102: Due to the small number of samples in the training data set, the training data set is randomly cropped, horizontally flipped and other operations are performed to enhance the data set, and the weight transfer learning method is used to prevent overfitting;
[0044] 103: Use the parallel convolution block proposed by the present invention on the last level of the ResNet34 model, such as image 3 As shown, instead of the original convolution block, the salient feature extraction ability of the network is enhanced;
[0045] Wherein, the ResNet34 model is well known to those skilled in the art, and will not be described in detail in this embodiment of the present invention.
[0046] 104: Input the training samples, the unif...
Embodiment 2
[0059] The following is combined with specific examples, calculation formulas, Figure 2-Figure 3 , for further introduction, see the description below:
[0060] Such as figure 2 As shown, the high-level salient features of the present invention are extracted by parallel convolutional blocks, and the fusion of inter-layer information is performed using bilinear pooling. suppose and is the high-level feature matrix extracted by parallel convolution blocks. The model proposed in this paper can be expressed by the following formula:
[0061] o bp =σ(N(β(X,Y))+N(β(X,Z))+N(β(Y,Z))) (1)
[0062] Among them, O bp Represents the final output of bilinear pairwise interaction between layers, namely figure 2 The corresponding output results, σ represents the softmax function, N represents the normalization operation, β represents the bilinear operation, R represents the dimensional space representation, h x Represents the length of each feature map of the first layer of the...
Embodiment 3
[0080] Combine below Figure 5-Figure 7 , Table 1-Table 2, the scheme in embodiment 1 and 2 is further introduced, see the following description for details:
[0081] Using the method of the present invention, on the three international standard fine-grained image libraries, compared with the unimproved convolution block, the heat maps on different channels are as follows: Figure 5 Shown: (a), (b), (c), (d) are the results of the CUB-200-2011 dataset; (e), (f) are the results of the FGVC-Aircraft dataset; (g), ( h) is the result of the Stanford-Cars dataset, the upper row uses the unimproved convolution block, and the lower row uses the improved convolution block. It can be seen that using parallel convolutional blocks can fully extract the salient regions of the image to be classified.
[0082] Using the method of the present invention, on three kinds of international standard fine-grained image libraries, the training and testing accuracy, and the corresponding confusion ...
PUM
Login to View More Abstract
Description
Claims
Application Information
Login to View More