基于CNN的多标签图像分类：标签维度异常问题求助

• 标签存储逻辑错误：原代码中labels_data[folder] = folder_data将图像数据赋值给标签字典，完全偏离了标签的定义，标签应对应文件夹的类别索引。
• 标签生成逻辑错误：生成all_labels时使用data[:, 0]作为模板，data是每个文件夹的图像数组（形状为(114,56,56)），data[:,0]提取的是每张图像的第一行像素，形状为(114,56)，因此np.full_like生成的数组形状与模板一致，最终拼接后得到(2280,56)，而非多标签分类所需的(2280,20)独热编码格式。

import cv2
import os
import numpy as np

def read_image(image_path):
    # 读取图像
    image = cv2.imread(image_path)
    # 转为灰度图
    image_resized = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
    # 归一化像素值
    image_resized = image_resized.astype('float32') / 255.0
    # 强制统一图像尺寸为(56,56)
    if image_resized.shape != (56,56):
        image_resized = cv2.resize(image_resized, (56,56), interpolation=cv2.INTER_AREA)
    return image_resized

def process_folder(folder_path):
    images = []
    # 获取文件夹内所有图像文件
    image_files = sorted([f for f in os.listdir(folder_path) if os.path.isfile(os.path.join(folder_path, f))])
    for image_file in image_files:
        image_path = os.path.join(folder_path, image_file)
        image_processed = read_image(image_path)
        images.append(image_processed)
    return np.array(images)

def create_folder_dictionary(root_folder_path):
    folder_dictionary = {}
    # 存储文件夹到类别索引的映射（0-19）
    folder_label_map = {}

    folders = sorted([f for f in os.listdir(root_folder_path) if os.path.isdir(os.path.join(root_folder_path, f))])
    for idx, folder in enumerate(folders):
        folder_path = os.path.join(root_folder_path, folder)
        folder_data = process_folder(folder_path)
        if folder_data is not None:
            folder_dictionary[folder] = folder_data
            folder_label_map[folder] = idx

    return folder_dictionary, folder_label_map


root_folder_path = r'C:\Users\sumit\Downloads\master thesis\ImageDataset'

result, folder_label_map = create_folder_dictionary(root_folder_path)

folders = list(result.keys())
folder_data_list = list(result.values())

# 统计总图像数
total_images = sum(len(images) for images in folder_data_list)
print(f"Total number of images: {total_images}")

# 拼接所有图像
all_images = np.concatenate(folder_data_list, axis=0)
print(f"All images shape: {all_images.shape}")

# 生成(2280,20)的独热编码标签
all_labels = []
for folder, data in zip(folders, folder_data_list):
    label_idx = folder_label_map[folder]
    # 为当前文件夹的每张图像生成独热编码
    one_hot_label = np.zeros((len(data), 20), dtype=np.float32)
    one_hot_label[:, label_idx] = 1.0
    all_labels.append(one_hot_label)

all_labels = np.concatenate(all_labels, axis=0)
print(f"Shape of all_labels: {all_labels.shape}")

• 移除错误的labels_data字典，改用folder_label_map存储文件夹与类别索引的对应关系。
• 生成标签时采用独热编码格式，每个标签是长度为20的数组，对应类别的位置设为1，其余为0，最终拼接后形状符合多标签分类的(2280,20)要求。
• 增加图像尺寸检查与强制统一，避免因原始图像尺寸不一致导致后续处理出错。

内容的提问来源于stack exchange，提问作者beschichtung346