Deep learning network processing device and method, and image processing unit

The invention relates to the technical field of image processing, and discloses a deep learning network processing method, which comprises the following steps: compressing to-be-processed image data into a data packet, uploading the compressed data packet to an image processing unit, and vectorizing the to-be-processed image data by using a data packet-based preprocessing mode by the image processing unit; performing multi-threshold selection of the image data vector by using an Otsu method based on an optimization algorithm to obtain an initial threshold of an image data vector histogram; carrying out image boundary enhancement on the image vector histogram by utilizing an image region boundary enhancement algorithm based on the histogram according to the selected initial threshold value; performing feature extraction on the enhanced image vector data by using a multi-channel sensing network; and performing image reconstruction processing on the extracted image features by using a multi-scale residual networ