def letterbox(img, new_shape=(640, 640), color=(114, 114, 114), scaleup=True):""":param img::param new_shape: (h,w):param color: pad填充颜色:param scaleup::return:"""shape = img.shape[:2]ratio = min(new_shape[0] / shape[0], new_shape[1] / shape[1])if not scaleup:ratio = min(ratio, 1.)new_w, new_h = int(round(shape[1] * ratio)), int(round(shape[0] * ratio))  # wxhdw, dh = new_shape[1] - new_w, new_shape[0] - new_hdw /= 2dh /= 2resized_img = cv2.resize(img, (new_w, new_h), interpolation=cv2.INTER_LINEAR)top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1))left, right = int(round(dw - 0.1)), int(round(dw + 0.1))# new_img = np.full((new_shape[0],new_shape[1],3),114, dtype=img.dtype)# new_img[top:top+new_h,left:left+new_w] = resized_img# cpu上copyborder更快new_img = cv2.copyMakeBorder(resized_img, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)  # add borderreturn new_img, ratio, (left, top)def letterbox_gpu(img, new_shape=(640, 640), color=(114, 114, 114), scaleup=True):""":param img::param new_shape: (h,w):param color: pad填充颜色:param scaleup::return:"""shape = img.shape[:2]ratio = min(new_shape[0] / shape[0], new_shape[1] / shape[1])if not scaleup:ratio = min(ratio, 1.)new_w, new_h = int(round(shape[1] * ratio)), int(round(shape[0] * ratio))  # wxhdw, dh = new_shape[1] - new_w, new_shape[0] - new_hdw /= 2dh /= 2top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1))left, right = int(round(dw - 0.1)), int(round(dw + 0.1))resized_img = zoom(img,(ratio,ratio,1),order=1)new_img = cp.full((new_shape[0],new_shape[1],3),114, dtype=resized_img.dtype)new_img[top:top+new_h,left:left+new_w] = resized_img# gpu上pad更慢# new_img = cp.pad(resized_img, ((top, bottom), (left, right), (0, 0)), mode='constant', constant_values=((114,114),(114,114),(114,114)))return new_img, ratio, (left, top)

测试图片1280x720，单纯测试letterbox，gpu版本要快0.24ms，但是gpu版本需要先将1280x720的原图拷贝到gpu，而cpu版本只需要拷贝640x640的结果图片到gpu，gpu拷贝这一步的耗时源超cpu版本，整体耗时基本相同。
结论：
在cpu上进行letterbox操作，在gpu上做归一化