Upload OmniParser/util/utils.py with huggingface_hub

OmniParser/util/utils.py

@@ -0,0 +1,628 @@
+# from ultralytics import YOLO
+import os
+import io
+import base64
+import time
+from PIL import Image, ImageDraw, ImageFont
+import json
+import requests
+# utility function
+import os
+from openai import AzureOpenAI
+
+import json
+import sys
+import os
+import cv2
+import numpy as np
+# %matplotlib inline
+from matplotlib import pyplot as plt
+import easyocr
+from paddleocr import PaddleOCR
+reader = easyocr.Reader(['en'])
+paddle_ocr = PaddleOCR(lang='en', use_angle_cls=False)
+import time
+import base64
+
+import os
+import ast
+import torch
+from typing import Tuple, List, Union
+from torchvision.ops import box_convert
+import re
+from torchvision.transforms import ToPILImage
+import supervision as sv
+import torchvision.transforms as T
+from util.box_annotator import BoxAnnotator 
+
+
+def get_caption_model_processor(model_name, model_name_or_path="Salesforce/blip2-opt-2.7b", device=None):
+    if not device:
+        device = "cuda" if torch.cuda.is_available() else "cpu"
+    if model_name == "blip2":
+        from transformers import Blip2Processor, Blip2ForConditionalGeneration
+        processor = Blip2Processor.from_pretrained("Salesforce/blip2-opt-2.7b")
+        if device == 'cpu':
+            model = Blip2ForConditionalGeneration.from_pretrained(
+            model_name_or_path, device_map=None, torch_dtype=torch.float32
+        ) 
+        else:
+            model = Blip2ForConditionalGeneration.from_pretrained(
+            model_name_or_path, device_map=None, torch_dtype=torch.float16
+        ).to(device)
+    else:
+        raise ValueError(f"Model {model_name} not supported. Only 'blip2' is available.")
+    return {'model': model.to(device), 'processor': processor}
+
+
+def get_yolo_model(model_path):
+    from ultralytics import YOLO
+    # Load the model.
+    model = YOLO(model_path)
+    return model
+
+
+@torch.inference_mode()
+def get_parsed_content_icon(filtered_boxes, starting_idx, image_source, caption_model_processor, prompt=None, batch_size=128):
+    # Number of samples per batch, --> 128 roughly takes 4 GB of GPU memory for blip2 model
+    to_pil = ToPILImage()
+    if starting_idx:
+        non_ocr_boxes = filtered_boxes[starting_idx:]
+    else:
+        non_ocr_boxes = filtered_boxes
+    croped_pil_image = []
+    for i, coord in enumerate(non_ocr_boxes):
+        try:
+            xmin, xmax = int(coord[0]*image_source.shape[1]), int(coord[2]*image_source.shape[1])
+            ymin, ymax = int(coord[1]*image_source.shape[0]), int(coord[3]*image_source.shape[0])
+            cropped_image = image_source[ymin:ymax, xmin:xmax, :]
+            cropped_image = cv2.resize(cropped_image, (64, 64))
+            croped_pil_image.append(to_pil(cropped_image))
+        except Exception as e:
+            print(f"[WARNING] Failed to crop image at {i}: {str(e)}")
+            continue
+
+    model, processor = caption_model_processor['model'], caption_model_processor['processor']
+    if not prompt:
+        prompt = "The image shows"
+    
+    # Reduce batch size on CPU to prevent crashes
+    device = model.device
+    if device.type == 'cpu':
+        batch_size = max(1, min(4, batch_size))  # Use max 4 on CPU
+        print(f"[Caption] Running on CPU, reducing batch_size to {batch_size}")
+    
+    print(f"[Caption] Processing {len(croped_pil_image)} cropped images with batch_size={batch_size}")
+    generated_texts = []
+    try:
+        for i in range(0, len(croped_pil_image), batch_size):
+            start = time.time()
+            batch = croped_pil_image[i:i+batch_size]
+            print(f"[Caption] Batch {i//batch_size + 1}: processing {len(batch)} images...")
+            
+            try:
+                t1 = time.time()
+                if model.device.type == 'cuda':
+                    inputs = processor(images=batch, text=[prompt]*len(batch), return_tensors="pt", do_resize=False).to(device=device, dtype=torch.float16)
+                else:
+                    inputs = processor(images=batch, text=[prompt]*len(batch), return_tensors="pt").to(device=device)
+                print(f"[Caption] Inputs prepared in {time.time()-t1:.2f}s")
+                
+                t2 = time.time()
+                generated_ids = model.generate(**inputs, max_length=100, num_beams=5, no_repeat_ngram_size=2, early_stopping=True, num_return_sequences=1)
+                print(f"[Caption] Generation done in {time.time()-t2:.2f}s")
+                
+                generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)
+                generated_text = [gen.strip() for gen in generated_text]
+                generated_texts.extend(generated_text)
+                print(f"[Caption] Batch complete in {time.time()-start:.2f}s")
+            except Exception as e:
+                print(f"[ERROR] Batch generation failed: {str(e)}")
+                import traceback
+                print(traceback.format_exc())
+                # Add placeholder captions for this batch instead of crashing
+                for j in range(len(batch)):
+                    generated_texts.append(f"Item {len(generated_texts)+1}")
+                print(f"[WARNING] Using placeholder captions for this batch")
+    except Exception as e:
+        print(f"[ERROR] Caption processing failed: {str(e)}")
+        import traceback
+        print(traceback.format_exc())
+        # Return placeholder captions if generation fails
+        generated_texts = [f"Item {i+1}" for i in range(len(croped_pil_image))]
+    
+    return generated_texts
+
+
+
+def get_parsed_content_icon_phi3v(filtered_boxes, ocr_bbox, image_source, caption_model_processor):
+    to_pil = ToPILImage()
+    if ocr_bbox:
+        non_ocr_boxes = filtered_boxes[len(ocr_bbox):]
+    else:
+        non_ocr_boxes = filtered_boxes
+    croped_pil_image = []
+    for i, coord in enumerate(non_ocr_boxes):
+        xmin, xmax = int(coord[0]*image_source.shape[1]), int(coord[2]*image_source.shape[1])
+        ymin, ymax = int(coord[1]*image_source.shape[0]), int(coord[3]*image_source.shape[0])
+        cropped_image = image_source[ymin:ymax, xmin:xmax, :]
+        croped_pil_image.append(to_pil(cropped_image))
+
+    model, processor = caption_model_processor['model'], caption_model_processor['processor']
+    device = model.device
+    messages = [{"role": "user", "content": "<|image_1|>\ndescribe the icon in one sentence"}] 
+    prompt = processor.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+
+    batch_size = 5  # Number of samples per batch
+    generated_texts = []
+
+    for i in range(0, len(croped_pil_image), batch_size):
+        images = croped_pil_image[i:i+batch_size]
+        image_inputs = [processor.image_processor(x, return_tensors="pt") for x in images]
+        inputs ={'input_ids': [], 'attention_mask': [], 'pixel_values': [], 'image_sizes': []}
+        texts = [prompt] * len(images)
+        for i, txt in enumerate(texts):
+            input = processor._convert_images_texts_to_inputs(image_inputs[i], txt, return_tensors="pt")
+            inputs['input_ids'].append(input['input_ids'])
+            inputs['attention_mask'].append(input['attention_mask'])
+            inputs['pixel_values'].append(input['pixel_values'])
+            inputs['image_sizes'].append(input['image_sizes'])
+        max_len = max([x.shape[1] for x in inputs['input_ids']])
+        for i, v in enumerate(inputs['input_ids']):
+            inputs['input_ids'][i] = torch.cat([processor.tokenizer.pad_token_id * torch.ones(1, max_len - v.shape[1], dtype=torch.long), v], dim=1)
+            inputs['attention_mask'][i] = torch.cat([torch.zeros(1, max_len - v.shape[1], dtype=torch.long), inputs['attention_mask'][i]], dim=1)
+        inputs_cat = {k: torch.concatenate(v).to(device) for k, v in inputs.items()}
+
+        generation_args = { 
+            "max_new_tokens": 25, 
+            "temperature": 0.01, 
+            "do_sample": False, 
+        } 
+        generate_ids = model.generate(**inputs_cat, eos_token_id=processor.tokenizer.eos_token_id, **generation_args) 
+        # # remove input tokens 
+        generate_ids = generate_ids[:, inputs_cat['input_ids'].shape[1]:]
+        response = processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)
+        response = [res.strip('\n').strip() for res in response]
+        generated_texts.extend(response)
+
+    return generated_texts
+
+def remove_overlap(boxes, iou_threshold, ocr_bbox=None):
+    assert ocr_bbox is None or isinstance(ocr_bbox, List)
+
+    def box_area(box):
+        return (box[2] - box[0]) * (box[3] - box[1])
+
+    def intersection_area(box1, box2):
+        x1 = max(box1[0], box2[0])
+        y1 = max(box1[1], box2[1])
+        x2 = min(box1[2], box2[2])
+        y2 = min(box1[3], box2[3])
+        return max(0, x2 - x1) * max(0, y2 - y1)
+
+    def IoU(box1, box2):
+        intersection = intersection_area(box1, box2)
+        union = box_area(box1) + box_area(box2) - intersection + 1e-6
+        if box_area(box1) > 0 and box_area(box2) > 0:
+            ratio1 = intersection / box_area(box1)
+            ratio2 = intersection / box_area(box2)
+        else:
+            ratio1, ratio2 = 0, 0
+        return max(intersection / union, ratio1, ratio2)
+
+    def is_inside(box1, box2):
+        # return box1[0] >= box2[0] and box1[1] >= box2[1] and box1[2] <= box2[2] and box1[3] <= box2[3]
+        intersection = intersection_area(box1, box2)
+        ratio1 = intersection / box_area(box1)
+        return ratio1 > 0.95
+
+    boxes = boxes.tolist()
+    filtered_boxes = []
+    if ocr_bbox:
+        filtered_boxes.extend(ocr_bbox)
+    # print('ocr_bbox!!!', ocr_bbox)
+    for i, box1 in enumerate(boxes):
+        # if not any(IoU(box1, box2) > iou_threshold and box_area(box1) > box_area(box2) for j, box2 in enumerate(boxes) if i != j):
+        is_valid_box = True
+        for j, box2 in enumerate(boxes):
+            # keep the smaller box
+            if i != j and IoU(box1, box2) > iou_threshold and box_area(box1) > box_area(box2):
+                is_valid_box = False
+                break
+        if is_valid_box:
+            # add the following 2 lines to include ocr bbox
+            if ocr_bbox:
+                # only add the box if it does not overlap with any ocr bbox
+                if not any(IoU(box1, box3) > iou_threshold and not is_inside(box1, box3) for k, box3 in enumerate(ocr_bbox)):
+                    filtered_boxes.append(box1)
+            else:
+                filtered_boxes.append(box1)
+    return torch.tensor(filtered_boxes)
+
+
+def remove_overlap_new(boxes, iou_threshold, ocr_bbox=None):
+    '''
+    ocr_bbox format: [{'type': 'text', 'bbox':[x,y], 'interactivity':False, 'content':str }, ...]
+    boxes format: [{'type': 'icon', 'bbox':[x,y], 'interactivity':True, 'content':None }, ...]
+
+    '''
+    assert ocr_bbox is None or isinstance(ocr_bbox, List)
+
+    def box_area(box):
+        return (box[2] - box[0]) * (box[3] - box[1])
+
+    def intersection_area(box1, box2):
+        x1 = max(box1[0], box2[0])
+        y1 = max(box1[1], box2[1])
+        x2 = min(box1[2], box2[2])
+        y2 = min(box1[3], box2[3])
+        return max(0, x2 - x1) * max(0, y2 - y1)
+
+    def IoU(box1, box2):
+        intersection = intersection_area(box1, box2)
+        union = box_area(box1) + box_area(box2) - intersection + 1e-6
+        if box_area(box1) > 0 and box_area(box2) > 0:
+            ratio1 = intersection / box_area(box1)
+            ratio2 = intersection / box_area(box2)
+        else:
+            ratio1, ratio2 = 0, 0
+        return max(intersection / union, ratio1, ratio2)
+
+    def is_inside(box1, box2):
+        # return box1[0] >= box2[0] and box1[1] >= box2[1] and box1[2] <= box2[2] and box1[3] <= box2[3]
+        intersection = intersection_area(box1, box2)
+        ratio1 = intersection / box_area(box1)
+        return ratio1 > 0.80
+
+    # boxes = boxes.tolist()
+    filtered_boxes = []
+    if ocr_bbox:
+        filtered_boxes.extend(ocr_bbox)
+    # print('ocr_bbox!!!', ocr_bbox)
+    for i, box1_elem in enumerate(boxes):
+        box1 = box1_elem['bbox']
+        is_valid_box = True
+        for j, box2_elem in enumerate(boxes):
+            # keep the smaller box
+            box2 = box2_elem['bbox']
+            if i != j and IoU(box1, box2) > iou_threshold and box_area(box1) > box_area(box2):
+                is_valid_box = False
+                break
+        if is_valid_box:
+            if ocr_bbox:
+                # keep yolo boxes + prioritize ocr label
+                box_added = False
+                ocr_labels = ''
+                for box3_elem in ocr_bbox:
+                    if not box_added:
+                        box3 = box3_elem['bbox']
+                        if is_inside(box3, box1): # ocr inside icon
+                            # box_added = True
+                            # delete the box3_elem from ocr_bbox
+                            try:
+                                # gather all ocr labels
+                                ocr_labels += box3_elem['content'] + ' '
+                                filtered_boxes.remove(box3_elem)
+                            except:
+                                continue
+                            # break
+                        elif is_inside(box1, box3): # icon inside ocr, don't added this icon box, no need to check other ocr bbox bc no overlap between ocr bbox, icon can only be in one ocr box
+                            box_added = True
+                            break
+                        else:
+                            continue
+                if not box_added:
+                    if ocr_labels:
+                        filtered_boxes.append({'type': 'icon', 'bbox': box1_elem['bbox'], 'interactivity': True, 'content': ocr_labels, 'source':'box_yolo_content_ocr'})
+                    else:
+                        filtered_boxes.append({'type': 'icon', 'bbox': box1_elem['bbox'], 'interactivity': True, 'content': None, 'source':'box_yolo_content_yolo'})
+            else:
+                filtered_boxes.append(box1)
+    return filtered_boxes # torch.tensor(filtered_boxes)
+
+
+def load_image(image_path: str) -> Tuple[np.array, torch.Tensor]:
+    transform = T.Compose(
+        [
+            T.RandomResize([800], max_size=1333),
+            T.ToTensor(),
+            T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
+        ]
+    )
+    image_source = Image.open(image_path).convert("RGB")
+    image = np.asarray(image_source)
+    image_transformed, _ = transform(image_source, None)
+    return image, image_transformed
+
+
+def annotate(image_source: np.ndarray, boxes: torch.Tensor, logits: torch.Tensor, phrases: List[str], text_scale: float, 
+             text_padding=5, text_thickness=2, thickness=3) -> np.ndarray:
+    """    
+    This function annotates an image with bounding boxes and labels.
+
+    Parameters:
+    image_source (np.ndarray): The source image to be annotated.
+    boxes (torch.Tensor): A tensor containing bounding box coordinates. in cxcywh format, pixel scale
+    logits (torch.Tensor): A tensor containing confidence scores for each bounding box.
+    phrases (List[str]): A list of labels for each bounding box.
+    text_scale (float): The scale of the text to be displayed. 0.8 for mobile/web, 0.3 for desktop # 0.4 for mind2web
+
+    Returns:
+    np.ndarray: The annotated image.
+    """
+    h, w, _ = image_source.shape
+    boxes = boxes * torch.Tensor([w, h, w, h])
+    xyxy = box_convert(boxes=boxes, in_fmt="cxcywh", out_fmt="xyxy").numpy()
+    xywh = box_convert(boxes=boxes, in_fmt="cxcywh", out_fmt="xywh").numpy()
+    detections = sv.Detections(xyxy=xyxy)
+
+    labels = [f"{phrase}" for phrase in range(boxes.shape[0])]
+
+    box_annotator = BoxAnnotator(text_scale=text_scale, text_padding=text_padding,text_thickness=text_thickness,thickness=thickness) # 0.8 for mobile/web, 0.3 for desktop # 0.4 for mind2web
+    annotated_frame = image_source.copy()
+    annotated_frame = box_annotator.annotate(scene=annotated_frame, detections=detections, labels=labels, image_size=(w,h))
+
+    label_coordinates = {f"{phrase}": v for phrase, v in zip(phrases, xywh)}
+    return annotated_frame, label_coordinates
+
+
+def predict(model, image, caption, box_threshold, text_threshold):
+    """ Use huggingface model to replace the original model
+    """
+    model, processor = model['model'], model['processor']
+    device = model.device
+
+    inputs = processor(images=image, text=caption, return_tensors="pt").to(device)
+    with torch.no_grad():
+        outputs = model(**inputs)
+
+    results = processor.post_process_grounded_object_detection(
+        outputs,
+        inputs.input_ids,
+        box_threshold=box_threshold, # 0.4,
+        text_threshold=text_threshold, # 0.3,
+        target_sizes=[image.size[::-1]]
+    )[0]
+    boxes, logits, phrases = results["boxes"], results["scores"], results["labels"]
+    return boxes, logits, phrases
+
+
+def predict_yolo(model, image, box_threshold, imgsz, scale_img, iou_threshold=0.7):
+    """ Use huggingface model to replace the original model
+    """
+    # model = model['model']
+    if scale_img:
+        result = model.predict(
+        source=image,
+        conf=box_threshold,
+        imgsz=imgsz,
+        iou=iou_threshold, # default 0.7
+        )
+    else:
+        result = model.predict(
+        source=image,
+        conf=box_threshold,
+        iou=iou_threshold, # default 0.7
+        )
+    boxes = result[0].boxes.xyxy#.tolist() # in pixel space
+    conf = result[0].boxes.conf
+    phrases = [str(i) for i in range(len(boxes))]
+
+    return boxes, conf, phrases
+
+def int_box_area(box, w, h):
+    x1, y1, x2, y2 = box
+    int_box = [int(x1*w), int(y1*h), int(x2*w), int(y2*h)]
+    area = (int_box[2] - int_box[0]) * (int_box[3] - int_box[1])
+    return area
+
+def get_som_labeled_img(image_source: Union[str, Image.Image], model=None, BOX_TRESHOLD=0.01, output_coord_in_ratio=False, ocr_bbox=None, text_scale=0.4, text_padding=5, draw_bbox_config=None, caption_model_processor=None, ocr_text=[], use_local_semantics=True, iou_threshold=0.9,prompt=None, scale_img=False, imgsz=None, batch_size=128, save_cropped_images=False, cropped_images_dir='cropped_images'):
+    """Process either an image path or Image object
+    
+    Args:
+        image_source: Either a file path (str) or PIL Image object
+        ...
+    """
+    if isinstance(image_source, str):
+        image_source = Image.open(image_source)
+    image_source = image_source.convert("RGB") # for CLIP
+    w, h = image_source.size
+    if not imgsz:
+        imgsz = (h, w)
+    # print('image size:', w, h)
+    xyxy, logits, phrases = predict_yolo(model=model, image=image_source, box_threshold=BOX_TRESHOLD, imgsz=imgsz, scale_img=scale_img, iou_threshold=0.1)
+    xyxy = xyxy / torch.Tensor([w, h, w, h]).to(xyxy.device)
+    image_source = np.asarray(image_source)
+    phrases = [str(i) for i in range(len(phrases))]
+
+    # annotate the image with labels
+    if ocr_bbox:
+        ocr_bbox = torch.tensor(ocr_bbox) / torch.Tensor([w, h, w, h])
+        ocr_bbox=ocr_bbox.tolist()
+    else:
+        print('no ocr bbox!!!')
+        ocr_bbox = None
+
+    if ocr_bbox is not None:
+        ocr_bbox_elem = [{'type': 'text', 'bbox':box, 'interactivity':False, 'content':txt, 'source': 'box_ocr_content_ocr'} for box, txt in zip(ocr_bbox, ocr_text) if int_box_area(box, w, h) > 0]
+    else:
+        ocr_bbox_elem = []
+    xyxy_elem = [{'type': 'icon', 'bbox':box, 'interactivity':True, 'content':None} for box in xyxy.tolist() if int_box_area(box, w, h) > 0]
+    filtered_boxes = remove_overlap_new(boxes=xyxy_elem, iou_threshold=iou_threshold, ocr_bbox=ocr_bbox_elem)
+    
+    # sort the filtered_boxes so that the one with 'content': None is at the end, and get the index of the first 'content': None
+    filtered_boxes_elem = sorted(filtered_boxes, key=lambda x: x['content'] is None)
+    # get the index of the first 'content': None
+    starting_idx = next((i for i, box in enumerate(filtered_boxes_elem) if box['content'] is None), -1)
+    filtered_boxes = torch.tensor([box['bbox'] for box in filtered_boxes_elem])
+    print('len(filtered_boxes):', len(filtered_boxes), starting_idx)
+
+    # get parsed icon local semantics
+    time1 = time.time()
+    if use_local_semantics and caption_model_processor is not None:
+        caption_model = caption_model_processor['model']
+        try:
+            print("[SOM] Starting caption generation...")
+            if 'phi3_v' in caption_model.config.model_type: 
+                print("[SOM] Using Phi3V model")
+                parsed_content_icon = get_parsed_content_icon_phi3v(filtered_boxes, ocr_bbox, image_source, caption_model_processor)
+            else:
+                print("[SOM] Using Caption model")
+                parsed_content_icon = get_parsed_content_icon(filtered_boxes, starting_idx, image_source, caption_model_processor, prompt=prompt,batch_size=batch_size)
+            print("[SOM] Caption generation complete")
+        except Exception as e:
+            print(f"[ERROR] Caption generation failed: {str(e)}")
+            import traceback
+            print(traceback.format_exc())
+            # Use placeholder captions if generation fails
+            num_icons = len(filtered_boxes) - (starting_idx if starting_idx > 0 else 0)
+            parsed_content_icon = [f"Icon {i}" for i in range(num_icons)]
+    elif use_local_semantics:
+        # Use OCR text if no caption model is available
+        print("[SOM] Caption model disabled, using OCR text only")
+        
+        # Save cropped images if requested
+        if save_cropped_images:
+            print(f"[SOM] Saving cropped UI images to {cropped_images_dir}...")
+            from pathlib import Path
+            Path(cropped_images_dir).mkdir(parents=True, exist_ok=True)
+            saved_count = 0
+            for i, bbox in enumerate(filtered_boxes):
+                try:
+                    # Convert from ratio coordinates to pixel coordinates
+                    xmin = int(bbox[0] * image_source.shape[1])
+                    xmax = int(bbox[2] * image_source.shape[1])
+                    ymin = int(bbox[1] * image_source.shape[0])
+                    ymax = int(bbox[3] * image_source.shape[0])
+                    
+                    # Ensure coordinates are within bounds
+                    xmin = max(0, xmin)
+                    ymin = max(0, ymin)
+                    xmax = min(image_source.shape[1], xmax)
+                    ymax = min(image_source.shape[0], ymax)
+                    
+                    # Crop the image
+                    cropped_image = image_source[ymin:ymax, xmin:xmax, :]
+                    
+                    # Save the cropped image
+                    pil_image = Image.fromarray(cropped_image)
+                    save_path = os.path.join(cropped_images_dir, f"crop_{i:04d}.png")
+                    pil_image.save(save_path)
+                    saved_count += 1
+                except Exception as e:
+                    print(f"[WARNING] Failed to save crop {i}: {str(e)}")
+            print(f"[SOM] Saved {saved_count} cropped UI images")
+        
+        parsed_content_icon = []
+        for i, bbox in enumerate(filtered_boxes):
+            # Check if any OCR text intersects with this bbox
+            label = f"Icon {i}"
+            if ocr_text and ocr_bbox is not None and len(ocr_text) > 0:
+                for j, ocr_b in enumerate(ocr_bbox):
+                    # Simple intersection check
+                    if (bbox[0] < ocr_b[2] and bbox[2] > ocr_b[0] and 
+                        bbox[1] < ocr_b[3] and bbox[3] > ocr_b[1]):
+                        label = ocr_text[j]
+                        break
+            parsed_content_icon.append(label)
+        
+        ocr_text = [f"Text Box ID {i}: {txt}" for i, txt in enumerate(ocr_text)]
+        icon_start = len(ocr_text)
+        parsed_content_icon_ls = []
+        # fill the filtered_boxes_elem None content with parsed_content_icon in order
+        for i, box in enumerate(filtered_boxes_elem):
+            if box['content'] is None:
+                box['content'] = parsed_content_icon.pop(0)
+        for i, txt in enumerate(parsed_content_icon):
+            parsed_content_icon_ls.append(f"Icon Box ID {str(i+icon_start)}: {txt}")
+        parsed_content_merged = ocr_text + parsed_content_icon_ls
+    else:
+        ocr_text = [f"Text Box ID {i}: {txt}" for i, txt in enumerate(ocr_text)]
+        parsed_content_merged = ocr_text
+    print('time to get parsed content:', time.time()-time1)
+
+    if len(filtered_boxes) > 0:
+        filtered_boxes = box_convert(boxes=filtered_boxes, in_fmt="xyxy", out_fmt="cxcywh")
+    else:
+        # No boxes detected, return empty results
+        print('No boxes detected in the image')
+        annotated_frame = image_source.copy()
+        label_coordinates = []
+        pil_img = Image.fromarray(annotated_frame)
+        buffered = io.BytesIO()
+        pil_img.save(buffered, format="PNG")
+        buffered.seek(0)
+        pil_str = base64.b64encode(buffered.getvalue()).decode()
+        return pil_str, label_coordinates, []
+
+    phrases = [i for i in range(len(filtered_boxes))]
+    
+    # draw boxes
+    if draw_bbox_config:
+        annotated_frame, label_coordinates = annotate(image_source=image_source, boxes=filtered_boxes, logits=logits, phrases=phrases, **draw_bbox_config)
+    else:
+        annotated_frame, label_coordinates = annotate(image_source=image_source, boxes=filtered_boxes, logits=logits, phrases=phrases, text_scale=text_scale, text_padding=text_padding)
+    
+    pil_img = Image.fromarray(annotated_frame)
+    buffered = io.BytesIO()
+    pil_img.save(buffered, format="PNG")
+    encoded_image = base64.b64encode(buffered.getvalue()).decode('ascii')
+    if output_coord_in_ratio:
+        label_coordinates = {k: [v[0]/w, v[1]/h, v[2]/w, v[3]/h] for k, v in label_coordinates.items()}
+        assert w == annotated_frame.shape[1] and h == annotated_frame.shape[0]
+
+    return encoded_image, label_coordinates, filtered_boxes_elem
+
+
+def get_xywh(input):
+    x, y, w, h = input[0][0], input[0][1], input[2][0] - input[0][0], input[2][1] - input[0][1]
+    x, y, w, h = int(x), int(y), int(w), int(h)
+    return x, y, w, h
+
+def get_xyxy(input):
+    x, y, xp, yp = input[0][0], input[0][1], input[2][0], input[2][1]
+    x, y, xp, yp = int(x), int(y), int(xp), int(yp)
+    return x, y, xp, yp
+
+def get_xywh_yolo(input):
+    x, y, w, h = input[0], input[1], input[2] - input[0], input[3] - input[1]
+    x, y, w, h = int(x), int(y), int(w), int(h)
+    return x, y, w, h
+
+def check_ocr_box(image_source: Union[str, Image.Image], display_img = True, output_bb_format='xywh', goal_filtering=None, easyocr_args=None, use_paddleocr=False):
+    if isinstance(image_source, str):
+        image_source = Image.open(image_source)
+    if image_source.mode == 'RGBA':
+        # Convert RGBA to RGB to avoid alpha channel issues
+        image_source = image_source.convert('RGB')
+    image_np = np.array(image_source)
+    w, h = image_source.size
+    if use_paddleocr:
+        if easyocr_args is None:
+            text_threshold = 0.5
+        else:
+            text_threshold = easyocr_args['text_threshold']
+        result = paddle_ocr.ocr(image_np, cls=False)[0]
+        coord = [item[0] for item in result if item[1][1] > text_threshold]
+        text = [item[1][0] for item in result if item[1][1] > text_threshold]
+    else:  # EasyOCR
+        if easyocr_args is None:
+            easyocr_args = {}
+        result = reader.readtext(image_np, **easyocr_args)
+        coord = [item[0] for item in result]
+        text = [item[1] for item in result]
+    if display_img:
+        opencv_img = cv2.cvtColor(image_np, cv2.COLOR_RGB2BGR)
+        bb = []
+        for item in coord:
+            x, y, a, b = get_xywh(item)
+            bb.append((x, y, a, b))
+            cv2.rectangle(opencv_img, (x, y), (x+a, y+b), (0, 255, 0), 2)
+        #  matplotlib expects RGB
+        plt.imshow(cv2.cvtColor(opencv_img, cv2.COLOR_BGR2RGB))
+    else:
+        if output_bb_format == 'xywh':
+            bb = [get_xywh(item) for item in coord]
+        elif output_bb_format == 'xyxy':
+            bb = [get_xyxy(item) for item in coord]
+    return (text, bb), goal_filtering