In-depth learning image recognition system applied in cellphone side and realizing method

Abstract

The invention discloses an in-depth learning image recognition system applied in a cellphone side and a realizing method. The realizing method includes: mapping to acquire a dendric category mapping relation according to concept dividing; on the basis of an original Inception-v3 network structure, adding a parallel connection branch at a dimensionality reduction position in a network to serve as input of a next layer of the network to acquire an improved Inception-v3 network structure; according to the improved Inception-v3 network structure, training on a setting category according to the dendric category mapping relation to acquire a base model; compressing the base model, and running the base model on the cellphone side to recognize an image, wherein compression at least includes one of parameter rarefaction, parameter quantization or reference sparse storage. Dendric image recognition based on concept category mapping is provided, and the original Inception-v3 model structure is improved in algorithm; a method for effectively compressing the model without having influence on accuracy of the improved model is provided, so that the model is ensured to effectively and stably run on the cellphone side.

Description

technical field [0001] The invention relates to mobile phone image recognition technology, in particular to a deep learning image recognition system and implementation method applied to mobile phones. Background technique [0002] Image recognition integrates many disciplines, including computer science and technology, physics, statistics, and neurobiology, and is widely used in geological exploration, image remote sensing, robot vision, biomedicine, and many other fields. Image recognition technology has many mature application cases on personal computers and embedded terminal devices, and with the continuous enhancement of mobile phone functions, this technology has gradually been applied to smart phones, but the weak processing power and low memory of smart phones themselves The limitations of the proposed test to the computationally complex recognition algorithm. [0003] At present, smart phones are developing very rapidly, basically integrating high-speed processing c...

AI Technical Summary

Problems solved by technology

However, compared with servers equipped with high-speed processing chips and large-capacity memory chips, smart phones still have the following disadvantages: lower image resolution, lack of dedicated graphics accelerators, no floating point units, especially low-end processors and Low-capacity memory, etc., limit complex calculations and large-scale calculations on smartphones

[0004] Specifically, there are many implementation methods in the existing mobile phone image recognition technology, such as single-level category image recognition on the mobile phone, but the disadvantage is that the number of categories is not rich enough, and the relationship between categories is not represented.

), but the disadvantage is that there are two places in the original structure that directly reduce the dimensionality of a Pooling layer, which is easy to cause a certain amount of information loss

For example, the model parameters are sparse, but the disadvantage is that a simple threshold is directly used to subtract the parameters, resulting in a certain loss of model accuracy.

Another example is the quantization of model parameters, but the disadvantages are: K-means clustering is easily affected by the number of parameters, and the robustness is not enough; the method of changing the quantization center is computationally intensive and the training speed is slow

For another example, the model parameters are stored sparsely, but the disadvantage is that the same bit size is used for all convolutional layers, and the storage efficiency is not high.

[0005] To sum up, most of today's mobile phone image recognition systems use deep models trained by deep learning for recognition. It is a problem to be solved to compress the model and reduce the amount of model storage and calculation to be suitable for the mobile phone.

Images