Image feature coding method based on DCT (Discrete Cosine Transform)

Abstract

The invention relates to an image feature coding method based on DCT (Discrete Cosine Transform), which comprises the following steps of: 1) extracting middle convolutional layer features of a convolutional neural network at an edge end; 2) the extracted features are tiled according to a channel sequence, and discrete cosine transform is carried out on the features; 3) performing K-means clustering quantization on the features after discrete cosine transform to obtain an index and a codebook; 4) performing arithmetic coding on the index and the codebook to obtain a bit stream and sending the bit stream to the cloud; and 5) at the cloud, performing decoding, inverse quantization, inverse discrete cosine transform and channelization on the encoded index and the encoded book to obtain a recovered feature vector. Compared with the prior art, the method has the advantages of effectively reducing the characteristic quantity, improving the transmission processing speed and the like.

Description

technical field [0001] The invention relates to the field of image compression coding, in particular to a DCT-based image feature coding method. Background technique [0002] As smart applications become more and more common in daily life, various data transmissions become more and more important, such as monitoring data in cities, detection data in industrial scenes, video data in smart driving, etc. As the amount of data captured and generated by devices such as sensors increases, the deep neural networks required to accomplish the required tasks may be too large or complex to be fully implemented within a single device. Therefore, devices at the edge of the network need to divide the deep neural network, compress the extracted intermediate data, and transmit the data to the backend through wireless channels, so that most of the calculations can be processed in the cloud. [0003] Compared with the size of the input or output data of the neural network, the size of the fe...

AI Technical Summary

Problems solved by technology

[0004] In the existing traditional image coding methods, the quantization module plays the main role of compression, and the uniform quantization method is adopted specifically. Even if the image features are quantized and coded, a large amount of data will still be generated. Therefore, in the scene where the transmission bandwidth is insufficient In this case, the transmission speed will be slow and the transmission efficiency will be seriously affected.

Images