I found that when I export model SAST to predict image, that made difference predict bounding box between inference model and train model

[cherry-pick] fix attn loss for ce

cp fix trt8 core bug

修复包导入错误

read_params应该从ocr_det而不是ocr_system中导入

add reference for locality_aware_nms

[cherry-pick 4552] add reference

[CP 4551 and 4555]add refer for some code

[fix infer] cp 

[cherry-pick] fix attn score

add nrtr reference

charry pick nrtr_postprocess and modify data type to adaptive

Update utility.py

Added other image formats that are already verified by imghdr.
Not sure about "pgm" format?