import random from dataclasses import dataclass from typing import Any, List, Dict, Optional, Union, Tuple import cv2 import torch import requests import numpy as np from PIL import Image import matplotlib.pyplot as plt from transformers import AutoModelForMaskGeneration, AutoProcessor, pipeline import gradio as gr @dataclass class BoundingBox: xmin: int ymin: int xmax: int ymax: int @property def xyxy(self) -> List[float]: return [self.xmin, self.ymin, self.xmax, self.ymax] @dataclass class DetectionResult: score: float label: str box: BoundingBox mask: Optional[np.array] = None @classmethod def from_dict(cls, detection_dict: Dict) -> 'DetectionResult': return cls(score=detection_dict['score'], label=detection_dict['label'], box=BoundingBox(xmin=detection_dict['box']['xmin'], ymin=detection_dict['box']['ymin'], xmax=detection_dict['box']['xmax'], ymax=detection_dict['box']['ymax'])) def annotate(image: Union[Image.Image, np.ndarray], detection_results: List[DetectionResult]) -> np.ndarray: image_cv2 = np.array(image) if isinstance(image, Image.Image) else image image_cv2 = cv2.cvtColor(image_cv2, cv2.COLOR_RGB2BGR) for detection in detection_results: label = detection.label score = detection.score box = detection.box mask = detection.mask color = np.random.randint(0, 256, size=3) cv2.rectangle(image_cv2, (box.xmin, box.ymin), (box.xmax, box.ymax), color.tolist(), 2) cv2.putText(image_cv2, f'{label}: {score:.2f}', (box.xmin, box.ymin - 10), cv2.FONT_HERSHEY_SIMPLEX, 0.5, color.tolist(), 2) if mask is not None: mask_uint8 = (mask * 255).astype(np.uint8) contours, _ = cv2.findContours(mask_uint8, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE) cv2.drawContours(image_cv2, contours, -1, color.tolist(), 2) return cv2.cvtColor(image_cv2, cv2.COLOR_BGR2RGB) def plot_detections(image: Union[Image.Image, np.ndarray], detections: List[DetectionResult]) -> np.ndarray: annotated_image = annotate(image, detections) return annotated_image def load_image(image: Union[str, Image.Image]) -> Image.Image: if isinstance(image, str) and image.startswith("http"): image = Image.open(requests.get(image, stream=True).raw).convert("RGB") elif isinstance(image, str): image = Image.open(image).convert("RGB") else: image = image.convert("RGB") return image def get_boxes(detection_results: List[DetectionResult]) -> List[List[List[float]]]: boxes = [] for result in detection_results: xyxy = result.box.xyxy boxes.append(xyxy) return [boxes] def refine_masks(masks: torch.BoolTensor, polygon_refinement: bool = False) -> List[np.ndarray]: masks = masks.cpu().float().permute(0, 2, 3, 1).mean(axis=-1).numpy().astype(np.uint8) masks = (masks > 0).astype(np.uint8) if polygon_refinement: for idx, mask in enumerate(masks): shape = mask.shape polygon = mask_to_polygon(mask) masks[idx] = polygon_to_mask(polygon, shape) return list(masks) def detect(image: Image.Image, labels: List[str], threshold: float = 0.3, detector_id: Optional[str] = None) -> List[Dict[str, Any]]: device = "cuda" if torch.cuda.is_available() else "cpu" detector_id = detector_id if detector_id else "IDEA-Research/grounding-dino-base" object_detector = pipeline(model=detector_id, task="zero-shot-object-detection", device=device) labels = [label if label.endswith(".") else label+"." for label in labels] results = object_detector(image, candidate_labels=labels, threshold=threshold) return [DetectionResult.from_dict(result) for result in results] def segment(image: Image.Image, detection_results: List[DetectionResult], polygon_refinement: bool = False, segmenter_id: Optional[str] = None) -> List[DetectionResult]: device = "cuda" if torch.cuda.is_available() else "cpu" segmenter_id = segmenter_id if segmenter_id else "martintmv/InsectSAM" segmentator = AutoModelForMaskGeneration.from_pretrained(segmenter_id).to(device) processor = AutoProcessor.from_pretrained(segmenter_id) boxes = get_boxes(detection_results) inputs = processor(images=image, input_boxes=boxes, return_tensors="pt").to(device) outputs = segmentator(**inputs) masks = processor.post_process_masks(masks=outputs.pred_masks, original_sizes=inputs.original_sizes, reshaped_input_sizes=inputs.reshaped_input_sizes)[0] masks = refine_masks(masks, polygon_refinement) for detection_result, mask in zip(detection_results, masks): detection_result.mask = mask return detection_results def grounded_segmentation(image: Union[Image.Image, str], labels: List[str], threshold: float = 0.3, polygon_refinement: bool = False, detector_id: Optional[str] = None, segmenter_id: Optional[str] = None) -> Tuple[np.ndarray, List[DetectionResult]]: image = load_image(image) detections = detect(image, labels, threshold, detector_id) detections = segment(image, detections, polygon_refinement, segmenter_id) return np.array(image), detections def extract_insect_masks(image: np.ndarray, detections: List[DetectionResult]) -> List[np.ndarray]: return [detection.mask for detection in detections if detection.mask is not None] def put_masks_on_yellow_background(image_shape: Tuple[int, int], masks: List[np.ndarray]) -> np.ndarray: yellow_background = np.full((image_shape[0], image_shape[1], 3), (0, 255, 255), dtype=np.uint8) for mask in masks: mask_rgb = cv2.cvtColor(mask, cv2.COLOR_GRAY2RGB) for c in range(3): yellow_background[:,:,c] = cv2.bitwise_or(yellow_background[:,:,c], mask_rgb[:,:,c]) return yellow_background def mask_to_min_max(mask: np.ndarray) -> Tuple[int, int, int, int]: y, x = np.where(mask) return x.min(), y.min(), x.max(), y.max() def extract_and_paste_insect(original_image: np.ndarray, detection: DetectionResult, background: np.ndarray) -> None: mask = detection.mask xmin, ymin, xmax, ymax = mask_to_min_max(mask) insect_crop = original_image[ymin:ymax, xmin:xmax] mask_crop = mask[ymin:ymax, xmin:xmax] insect = cv2.bitwise_and(insect_crop, insect_crop, mask=mask_crop) x_offset, y_offset = detection.box.xmin, detection.box.ymin x_end, y_end = x_offset + insect.shape[1], y_offset + insect.shape[0] inverse_mask = cv2.bitwise_not(mask_crop) bg_region = background[y_offset:y_end, x_offset:x_end] bg_ready = cv2.bitwise_and(bg_region, bg_region, mask=inverse_mask) combined = cv2.add(insect, bg_ready) background[y_offset:y_end, x_offset:x_end] = combined def create_yellow_background_with_insects(image: np.ndarray, detections: List[DetectionResult]) -> np.ndarray: yellow_background = np.full_like(image, (0, 255, 255), dtype=np.uint8) for detection in detections: if detection.mask is not None: extract_and_paste_insect(image, detection, yellow_background) return yellow_background def draw_classification_boxes(image_with_insects: np.ndarray, detections: List[DetectionResult]) -> np.ndarray: for detection in detections: label = detection.label score = detection.score box = detection.box color = np.random.randint(0, 256, size=3).tolist() cv2.rectangle(image_with_insects, (box.xmin, box.ymin), (box.xmax, box.ymax), color, 2) (text_width, text_height), baseline = cv2.getTextSize(f"{label}: {score:.2f}", cv2.FONT_HERSHEY_SIMPLEX, 0.5, 2) cv2.rectangle(image_with_insects, (box.xmin, box.ymin - text_height - baseline), (box.xmin + text_width, box.ymin), color, thickness=cv2.FILLED) cv2.putText(image_with_insects, f"{label}: {score:.2f}", (box.xmin, box.ymin - baseline), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 0), 2) return image_with_insects def process_image(image): labels = ["ant", "beetle", "butterfly", "caterpillar", "dragonfly"] original_image, detections = grounded_segmentation(image, labels, threshold=0.3, polygon_refinement=True) masked_image = plot_detections(original_image, detections) insect_masks = extract_insect_masks(original_image, detections) yellow_background_with_masks = put_masks_on_yellow_background(original_image.shape[:2], insect_masks) yellow_background_with_insects = create_yellow_background_with_insects(original_image, detections) yellow_background_with_boxes = draw_classification_boxes(yellow_background_with_insects, detections) return masked_image, yellow_background_with_masks, yellow_background_with_boxes gr.Interface( fn=process_image, inputs=gr.Image(type="pil"), outputs=[gr.Image(type="numpy"), gr.Image(type="numpy"), gr.Image(type="numpy")], title="Insect Detection and Masking" ).launch()