Method and device for object recognition
A target recognition and equipment technology, applied in the field of target recognition, can solve the problems of network loss of regularity and unfavorable parallelism, and achieve the effect of maintaining the network performance unchanged, increasing the degree of tailoring, and speeding up the training speed.
- Summary
- Abstract
- Description
- Claims
- Application Information
AI Technical Summary
Problems solved by technology
Method used
Image
Examples
Embodiment 1
[0043] The present embodiment provides a method of target recognition, such as figure 1 Shown, includes the following steps:
[0044] Step 101, acquiring image data for target recognition;
[0045] Step 102, the input image data for object recognition neural network model, to obtain a recognition result of the recognition target, wherein the preset input training samples using the neural network model is trained to give the neural network model in the pre- process set up to train the neural network model, it is determined whether the predetermined neural network model includes bulk normalized BN layer; if so, the pre-BN layer neural network model number of layers were cut, otherwise, the right said predetermined network layer neural network model weights crop.
[0046] In the above method, of the weight of the weight of the structure itself through a predetermined neural network model combined direct and indirect methods of sparsity structure of BN layer was cut, greatly accelerat...
specific Embodiment approach
[0051] If the neural network is determined when the predetermined current comprises a layer of BN, the current preset number of layers of the neural network model BN layer is cut. DETAILED DESCRIPTION follows:
[0052] As an optional embodiment, by adding the second term to the predetermined penalty BN layer neural network model, the preset number of layers of the neural network model BN layer is cut.
[0053] The second penalty term includes a second adjustment factor for adjusting the predetermined level and tailoring BN BN layer output value range of the output value of the neural network model;
[0054] To preset the BN layer neural network model, according to a second penalty term in the second adjustment coefficient, the output value of the BN layer in a predetermined neural network model is adjusted;
[0055] , The output value of the neural network model preset BN layer is adjusted by dynamically adjusting the penalty of the second term in the second adjustment coefficient...
Embodiment 2
[0073] Based on the same inventive concept, the embodiments provide an object recognition apparatus according to the present embodiment, as image 3 Shown, the apparatus 302 includes a processor 301 and a memory, wherein the memory 302 stores the executable program, the processor 301 implemented as an executable procedure when said program is executed:
[0074] Acquiring image data for target recognition;
[0075] The image data input to the neural network model target recognition, to obtain a recognition result of the recognition target, wherein the preset input training samples using the neural network model is trained to give the neural network, the neural network in the preset model training process, the neural network model to determine whether the preset batch includes a normalized BN layer; if so, the number of layers of pre-BN layer of the neural network model crop, otherwise, the default the right to the network layer of the neural network model of heavy crop.
[0076] As ...
PUM
Login to View More Abstract
Description
Claims
Application Information
Login to View More