From 29cadea1a49089f325d95c35d6b0ff5e6cc3500e Mon Sep 17 00:00:00 2001 From: Bairly <2652270566@qq.com> Date: Tue, 1 Apr 2025 19:46:25 +0800 Subject: [PATCH] add:covers_analyser --- covers_analyser.py | 181 +++++++++++++++++++-------------------------- 1 file changed, 77 insertions(+), 104 deletions(-) diff --git a/covers_analyser.py b/covers_analyser.py index 715a740..f50bed9 100644 --- a/covers_analyser.py +++ b/covers_analyser.py @@ -5,40 +5,17 @@ import requests from io import BytesIO from PIL import Image import os -from colorthief import ColorThief -import pytesseract from multiprocessing import Pool -from cnsenti import Sentiment -import pynlpir -from collections import defaultdict +from sklearn.cluster import MiniBatchKMeans +from tqdm import tqdm import warnings warnings.filterwarnings('ignore') -#设置OCR路径 -pytesseract.pytesseract.tesseract_cmd = r'D:Program files\Tesseract-OCR\tesseract.exe' - # ------------------图像处理-初始化配置 --------------------- # 人脸检测模型初始化 face_cascade = cv2.CascadeClassifier(cv2.data.haarcascades + 'haarcascade_frontalface_default.xml') -# 图像情感模型系数（基于IAPS数据集校准） -VALENCE_WEIGHTS = { - 'warm_ratio': 0.35, - 'brightness': 0.15, - 'symmetry': 0.20, - 'colorfulness': 0.30 -} - -AROUSAL_WEIGHTS = { - 'contrast': 0.40, - 'edge_density': 0.35, - 'saturation_std': 0.25 -} - -# 暖色调定义（HSV色相范围） -WARM_HUE_RANGE = (0, 60) # 红色到黄色 - # ------------------处理图像 --------------------- def get_image(url): """从URL获取图像并预处理""" @@ -51,56 +28,72 @@ def get_image(url): return None -def analyze_image(img): - """分析图像特征""" - if img is None: - return {} +def extract_color_palette(img_rgb, color_count=5): + """提取前N种主色及比例""" + # 使用k-means聚类提取主色 + pixels = img_rgb.reshape(-1, 3) + kmeans = MiniBatchKMeans(n_clusters=color_count, random_state=0) + labels = kmeans.fit_predict(pixels) - # 转换为HSV颜色空间 - hsv = cv2.cvtColor(img, cv2.COLOR_BGR2HSV) - h, s, v = cv2.split(hsv) + # 计算每种颜色的比例 + counts = np.bincount(labels) + total = counts.sum() + palette = [] + for i in range(color_count): + ratio = counts[i] / total + color = kmeans.cluster_centers_[i].astype(int) + palette.append((color, ratio)) - # 计算基础特征 - features = { - 'brightness': np.mean(v), - 'contrast': np.max(v) - np.min(v), - 'saturation_std': np.std(s), - 'colorfulness': np.std(h) + np.std(s) + np.std(v) - } - - # 暖色比例计算 - hue_mask = cv2.inRange(h, WARM_HUE_RANGE[0], WARM_HUE_RANGE[1]) - features['warm_ratio'] = np.count_nonzero(hue_mask) / (img.shape[0] * img.shape[1]) - - # 对称性计算 - mid = img.shape[1] // 2 - left = img[:, :mid] - right = cv2.flip(img[:, mid:], 1) - features['symmetry'] = cv2.matchTemplate(left, right, cv2.TM_CCOEFF_NORMED)[0][0] - - # 边缘密度 - edges = cv2.Canny(cv2.cvtColor(img, cv2.COLOR_BGR2GRAY), 100, 200) - features['edge_density'] = np.mean(edges) - - return features + return sorted(palette, key=lambda x: -x[1]) # 按比例降序排列 -def calculate_affect(features): - """计算情感效价和唤醒度""" - poslm = sum(features[k] * VALENCE_WEIGHTS[k] for k in VALENCE_WEIGHTS) - actlm = sum(features[k] * AROUSAL_WEIGHTS[k] for k in AROUSAL_WEIGHTS) +def classify_hsv_color(rgb_color): + """将RGB颜色分类为暖色/冷色/中性色""" + try: + # 确保输入是有效的RGB颜色值 + rgb_color = np.clip(rgb_color, 0, 255) + hsv = cv2.cvtColor(np.uint8([[rgb_color]]), cv2.COLOR_RGB2HSV)[0][0] + h, s, v = hsv[0], hsv[1] / 255.0, hsv[2] / 255.0 # 归一化 - # Sigmoid归一化 - return { - 'Poslm': 2 / (1 + np.exp(-poslm)) - 1, - 'Actlm': 2 / (1 + np.exp(-actlm)) - 1 - } + # 中性色判断（根据Palmer标准） + if s < 0.2 or v < 0.2: + return 'neutral' + + # 色相分类 + if (0 <= h < 90) or (270 <= h <= 360): + return 'warm' + return 'cool' + except Exception as e: + print(f"颜色越界: {str(e)}") + return 'neutral' # 出错时默认返回中性色 + +def determine_warm_tone(img_rgb): + """返回（暖色标签, 暖色比例, 冷色比例）""" + palette = extract_color_palette(img_rgb) + + warm_ratio, cool_ratio, neutral_ratio = 0.0, 0.0, 0.0 + for color, ratio in palette: + category = classify_hsv_color(color) + if category == 'warm': + warm_ratio += ratio + elif category == 'cool': + cool_ratio += ratio + else: + neutral_ratio += ratio + + return warm_ratio, cool_ratio, neutral_ratio def detect_human(img): - """检测人像""" gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) - faces = face_cascade.detectMultiScale(gray, 1.05, 3) + # 优化参数组合 + faces = face_cascade.detectMultiScale( + gray, + scaleFactor=1.02, # 减小缩放步长，增加检测粒度 + minNeighbors=5, # 提高邻居数要求，减少误检 + minSize=(50, 50), # 适配B站封面最小人脸尺寸 + flags=cv2.CASCADE_FIND_BIGGEST_OBJECT # 优先检测最大人脸 + ) return len(faces) > 0 @@ -110,24 +103,18 @@ def process_url(url): if img is None: return None - features = analyze_image(img) - affect = calculate_affect(features) - - # 主色分析 - color_thief = ColorThief(BytesIO(requests.get(url).content)) - dominant_color = color_thief.get_color(quality=1) - hsv_color = cv2.cvtColor(np.uint8([[dominant_color]]), cv2.COLOR_RGB2HSV)[0][0] - warm = 1 if WARM_HUE_RANGE[0] <= hsv_color[0] <= WARM_HUE_RANGE[1] else 0 - - # 人像检测 - has_human = detect_human(img) + # 颜色分析 + img_rgb = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) + warm_ratio, cool_ratio, neutral_ratio = determine_warm_tone(img_rgb) return { 'url': url, - **affect, - 'Warm': warm, - 'Portrait': int(has_human) + 'Portrait': int(detect_human(img)), + 'WarmRatio': round(warm_ratio, 3), + 'CoolRatio': round(cool_ratio, 3), + 'NeutralRatio': round(neutral_ratio, 3) } + except Exception as e: print(f"Error processing {url}: {str(e)}") return None @@ -136,41 +123,27 @@ def process_url(url): # 批量处理 def batch_process(urls, workers=4): with Pool(workers) as pool: - results = [res for res in pool.imap(process_url, urls) if res is not None] + results = [res for res in tqdm(pool.imap(process_url, urls), + total=len(urls), + desc="处理进度") if res is not None] return pd.DataFrame(results) # 使用示例 if __name__ == "__main__": # 读取URL列表 - input_csv = "data_all.csv" + input_csv = "data_all_first_ver.csv" #输出路径 os.makedirs('./result', exist_ok=True) output_csv = "result/analysis_results.csv" - ##完整运行 - # df = pd.read_csv(input_csv) - # urls = df['视频封面'].tolist() - # - # # 执行分析 - # result_df = batch_process(urls) - # - # # 合并原始数据 - # final_df = df.merge(result_df, left_on='视频封面') - # final_df.drop('url', axis=1).to_csv(output_csv, index=False) + #完整运行 + df = pd.read_csv(input_csv) + urls = df['视频封面URL'].tolist() - # 示例URL列表 - #小批量实验 - urls = [ - 'http://i0.hdslb.com/bfs/archive/393a8e961b704d43256fe7e6c89fee04df966e17.jpg', - 'http://i0.hdslb.com/bfs/archive/072e16a1237040941f15b1ed67a8d1ebe6f2e041.jpg', - 'http://i2.hdslb.com/bfs/archive/1c56b5bec767c604175983cc5926f5832baa9bb8.jpg', - 'http://i0.hdslb.com/bfs/archive/66384e53a15345a539ccbb2989442f1d960b9235.jpg', - 'http://i2.hdslb.com/bfs/archive/836b762456f0b4d65dd2c40fc4cd120107e46b88.jpg', - ] + # 执行分析 result_df = batch_process(urls) - result_df.to_csv("result/analysis_results.csv", index=False) + result_df.to_csv(output_csv, index=False, encoding='utf-8-sig') + print(f"成功处理 {len(result_df)}/{len(urls)} 张图片") - - print("分析完成！结果已保存至", output_csv) \ No newline at end of file