Automatic commodity classification method and system with interpretability
An automatic classification and explanatory technology, applied in neural learning methods, business, biological neural network models, etc., can solve problems such as classification, difficult and complex commodities, and commodity image recognition classification algorithms that are difficult to obtain interpretation, and achieve high classification accuracy , strong explanatory power, and the effect of saving debugging time
- Summary
- Abstract
- Description
- Claims
- Application Information
AI Technical Summary
Problems solved by technology
Method used
Image
Examples
specific Embodiment approach 1
[0020] Specific implementation mode 1. Combination figure 1 This embodiment will be described. An explanatory automatic classification method for commodities described in this embodiment, the method specifically includes the following steps:
[0021] Step 1. Design a network architecture consisting of a basic classifier and an interpretable level, wherein the basic classifier is a CNN network model, and the interpretable level is an induced decision tree;
[0022] After pre-training the CNN network model, the induced decision tree is constructed by loading the weights of the last fully connected layer in the pre-trained CNN network model;
[0023] Step 2. Use the marked commodity image data set to train the network architecture designed in step 1 until the classification accuracy of the network architecture on the marked data set reaches the set threshold and stop training;
[0024] Step 3: After inputting the image of the product to be classified into the trained network ar...
specific Embodiment approach 2
[0034] Specific embodiment 2: The difference between this embodiment and specific embodiment 1 is that the CNN network model uses the Resnet18 model as the basic construction, and each Resnet18 model includes a convolutional layer, a Pooling layer, four blocks and a full connection layer;
[0035] Among them, each block is composed of two build blocks, and each build block is composed of two small squares.
[0036] Other steps and parameters are the same as those in Embodiment 1.
specific Embodiment approach 3
[0037] Specific implementation mode three: combination figure 2 This embodiment will be described. The difference between this embodiment and the specific embodiment 1 or 2 is that the induced decision tree is constructed by loading the weight of the last fully connected layer in the pre-trained CNN network model, and the specific process is:
[0038] Step S1, load the weight of the last fully connected layer in the CNN network model obtained by pre-training;
[0039] Step S2, taking each row of loaded weights as a representative vector of a leaf node;
[0040] Step S3, using the representative average vector of each pair of leaf nodes as the representative vector of the parents;
[0041] Step S4, taking the representative average vectors of all leaf nodes as the representative vectors of the ancestors.
[0042] Other steps and parameters are the same as those in Embodiment 1 or Embodiment 2.
[0043] Build hierarchies on the weight space to obtain interpretable models wi...
PUM
Login to View More Abstract
Description
Claims
Application Information
Login to View More