LSTM-based two-dimensional image target class identification method

Abstract

The invention provides an LSTM-based two-dimensional image target class identification method comprising the following steps that S1, a two-dimensional image is converted into a one-dimensional vectoraccording to a ring series mode from outside to inside, the two-dimensional image is converted into the one-dimensional vector according to a row sequence series mode, and the two-dimensional image is converted into the one-dimensional vector according to a column sequence series mode; S2, the one-dimensional vectors obtained by the three conversion modes are all transmitted to an LSTM network tobe trained or identified, and then three results are outputted; and S3, the three results outputted by the LSTM network are fused and target identification is performed based on the final fusion result. The beneficial effects of the method are that different modes are applied to the two-dimensional image to obtain the one-dimensional vector so that the spatial organization relationship of the two-dimensional image can be fully utilized, the correct rate of identification can be obviously enhanced and the method has great interpretability.

Description

technical field [0001] The present invention relates to image recognition, in particular to an LSTM-based two-dimensional image target category recognition method. Background technique [0002] LSTM (Long Short Term Memory) deep network has important applications in object category recognition, but it is limited to the situation where the object is expressed as one-dimensional data. The main advantage of the LSTM network is that it can make full use of the associated information of the sequence in time sequence. figure 1 A schematic diagram of the overall structure of LSTM is given. In order to solve the problem of gradient disappearance in the traditional RNN network, that is, the perception of the later time nodes to the previous time nodes decreases. LSTM has designed a unique Cell structure to realize the memory function and maintain the perception ability of subsequent time nodes for previous time nodes. LSTM also designed a "Forget Gate" to selectively forget previous...

AI Technical Summary

Problems solved by technology

The one-dimensional vector obtained by such a single conversion method cannot fully reflect the spatial organization relationship of image pixels, that is, this method ignores other optional methods when matrix data is converted into one-dimensional data; and makes the two-dimensional topology of image data ( The distance relationship of different pixels in two-dimensional space) is not fully reflected, resulting in a low accuracy rate of discrimination

Images