import os from matplotlib import pyplot as plt import skimage from skimage import io as skimage_io import numpy as np import re def rescale_detection_box(boxes, image): h_img, w_img, _ = image.shape size = max(h_img, w_img) pad_h = size - h_img pad_w = size - w_img recovered_boxes = [] for box in boxes: cx, cy, w, h = box cx = cx * size cy = cy * size w = w * size h = h * size # if cx < 0 or cx > w_img or cy < 0 or cy > h_img: # continue; x1 = cx - w / 2 y1 = cy - h / 2 x2 = cx + w / 2 y2 = cy + h / 2 recovered_boxes.append((x1, y1, x2, y2)) return recovered_boxes def plot_boxes_on_image(image, text_queries, scores, boxes, labels, filename, score_threshold, output_dir): colors = ['red', 'green', 'blue', 'orange', 'purple', 'pink', 'cyan', 'magenta', 'lightblue', 'darkorange', 'darkgreen', 'darkred', 'lavender', 'brown', 'gray', 'black'] # 显示原始图片 plt.clf() plt.imshow(image) plt.axis('off') # 绘制边界框 for score, box, label in zip(scores, boxes, labels): if score < score_threshold: continue; x1, y1, x2, y2 = box # print(f"box coord: {[x1, y1, x2, y2]}") plt.plot( [x1, x2, x2, x1, x1], [y1, y1, y2, y2, y1], color=colors[label], linewidth=0.6, alpha=0.6 ) plt.text( x1, y2 + 0.015, f'{text_queries[label]}: {score:1.2f}', ha='left', va='top', color=colors[label], fontsize=6, bbox={'facecolor': 'white', 'edgecolor': colors[label], 'boxstyle': 'square,pad=.3', 'alpha': 0.5} ) # 保存图片到指定路径 OUTPUT_DIR output_path = os.path.join(output_dir, filename) plt.savefig(output_path, bbox_inches='tight', pad_inches=0.1, dpi=300) print(f"Image with boxes saved to {output_path}") def image_based_plot_boxes_on_image(image, text_queries, scores, boxes, filename,output_dir): colors = ['red', 'green', 'blue', 'orange', 'cyan', 'magenta', 'lightblue', 'darkorange', 'lavender'] plt.clf() plt.imshow(image) plt.axis('off') for score, box, text_query, color in zip(scores, boxes, text_queries, colors): x1, y1, x2, y2 = box plt.plot( [x1, x2, x2, x1, x1], [y1, y1, y2, y2, y1], color=color, linewidth=1 ) plt.text( x1, y2 + 0.015, f'{text_query}: {score:1.2f}', ha='left', va='top', color=color, fontsize=6, bbox={'facecolor': 'white', 'edgecolor': color, 'boxstyle': 'square,pad=.3','alpha': 0.5} ) output_path = os.path.join(output_dir, filename) plt.savefig(output_path, bbox_inches='tight', pad_inches=0.1, dpi=300) print(f"Image with boxes saved to {output_path}") def get_iou(bbox1, bbox2): # 分别解包 bbox1 和 bbox2 的坐标 x1_min, y1_min, x1_max, y1_max = bbox1 x2_min, y2_min, x2_max, y2_max = bbox2 # 计算交集的顶点坐标 inter_x_min = max(x1_min, x2_min) inter_y_min = max(y1_min, y2_min) inter_x_max = min(x1_max, x2_max) inter_y_max = min(y1_max, y2_max) # 计算交集的宽度和高度（确保为非负值） inter_width = max(0, inter_x_max - inter_x_min) inter_height = max(0, inter_y_max - inter_y_min) inter_area = inter_width * inter_height # 计算每个边界框的面积 area1 = (x1_max - x1_min) * (y1_max - y1_min) area2 = (x2_max - x2_min) * (y2_max - y2_min) # 计算并集面积 union_area = area1 + area2 - inter_area # 计算 IOU iou = inter_area / union_area if union_area > 0 else 0 return iou def read_images(image_dir): images = [] filenames = sorted(os.listdir(image_dir)) for filename in filenames: file_path = os.path.join(image_dir, filename) image_uint8 = skimage_io.imread(file_path) image = image_uint8.astype(np.float32) / 255.0 images.append(image) return images, filenames def preprocess_images(images, model_input_size): processed_images = [] for image in images: # Pad image to square h, w, d = image.shape size = max(h, w) image_padded = np.pad(image, ((0, size - h), (0, size - w), (0, 0)), constant_values=0.5,) # Resize image to fit model's input size image_resized = skimage.transform.resize( image_padded, (model_input_size, model_input_size), anti_aliasing=True, ) processed_images.append(image_resized) # Shape: (b, h, w, d) return np.array(processed_images, dtype=np.float32) def prepare_images(image_dir, model_input_size): filenames = sorted(os.listdir(image_dir)) images = [] for filename in filenames: file_path = os.path.join(image_dir, filename) image_uint8 = skimage_io.imread(file_path) image = image_uint8.astype(np.float32) / 255.0 # Pad image to square h, w, d = image.shape size = max(h, w) image_padded = np.pad( image, ((0, size - h), (0, size - w), (0, 0)), constant_values=0.5 ) # Resize image to fit model's input size image_resized = skimage.transform.resize( image_padded, (model_input_size, model_input_size), anti_aliasing=True, ) images.append(image_resized) # Shape: (b, h, w, d) return np.array(images, dtype=np.float32), filenames def plot_bbox_on_image(image, boxes, objectnesses, threshold, output_file): fig, ax = plt.subplots(1, 1, figsize=(8, 8)) ax.imshow(image, extent=(0, 1, 1, 0)) ax.set_axis_off() for i, (box, objectness) in enumerate(zip(boxes, objectnesses)): if objectness < threshold: continue index = i cx, cy, w, h = box ax.plot( [cx - w / 2, cx + w / 2, cx + w / 2, cx - w / 2, cx - w / 2], [cy - h / 2, cy - h / 2, cy + h / 2, cy + h / 2, cy - h / 2], color='lime', ) ax.text( cx - w / 2 + 0.015, cy + h / 2 - 0.015, f'Index {i}: {objectness:1.2f}', ha='left', va='bottom', color='black', bbox={ 'facecolor': 'white', 'edgecolor': 'lime', 'boxstyle': 'square,pad=.3', }, ) ax.set_xlim(0, 1) ax.set_ylim(1, 0) ax.set_title(f'Top objects by objectness') # 保存图片到指定路径 plt.savefig(output_file, bbox_inches='tight', dpi=300) plt.close() # 关闭图像以释放内存 print(f"结果图片已保存到: {output_file}") return index def top_object_index(objectnesses, threshold): for i, objectness in enumerate(objectnesses): if objectness < threshold: continue else: return i def boxes_filter(pred_bboxes, raw_bboxes, pred_scores, instances): # Step 1: Filter by pred_scores filtered_indices = [i for i, score in enumerate(pred_scores) if score >= 0.97] pred_bboxes = [pred_bboxes[i] for i in filtered_indices] raw_bboxes = [raw_bboxes[i] for i in filtered_indices] pred_scores = [pred_scores[i] for i in filtered_indices] instances = [instances[i] for i in filtered_indices] # Step 2: Filter by IoU keep_indices = set(range(len(pred_bboxes))) for i in range(len(pred_bboxes)): if i not in keep_indices: continue for j in range(i + 1, len(pred_bboxes)): if j not in keep_indices: continue iou = get_iou(pred_bboxes[i], pred_bboxes[j]) if iou > 0.9: if pred_scores[i] >= pred_scores[j]: keep_indices.discard(j) else: keep_indices.discard(i) pred_bboxes = [pred_bboxes[i] for i in sorted(keep_indices)] raw_bboxes = [raw_bboxes[i] for i in sorted(keep_indices)] pred_scores = [pred_scores[i] for i in sorted(keep_indices)] instances = [instances[i] for i in sorted(keep_indices)] # Step 3: Filter by duplicate instances instance_map = {} for i in range(len(instances)): instance = instances[i] if instance not in instance_map or pred_scores[i] > pred_scores[instance_map[instance]]: instance_map[instance] = i unique_indices = sorted(instance_map.values()) pred_bboxes = [pred_bboxes[i] for i in unique_indices] raw_bboxes = [raw_bboxes[i] for i in unique_indices] pred_scores = [pred_scores[i] for i in unique_indices] instances = [instances[i] for i in unique_indices] return pred_bboxes, raw_bboxes, pred_scores, instances def format_string(input_string: str) -> str: # 大写 转 小写 lowercased = input_string.lower() # 空格 转 下划线 transformed = re.sub(r"\s+", "_", lowercased) # \s+ 匹配一个或多个空白字符 return transformed