Python将图片写到音频频谱

向右看齐 2022-06-03 07:29 231阅读 0赞

用到的库：pillow、numpy、matplotlib

[源码GitHub][GitHub]

# -*- coding: utf-8 -*-
    
    import wave
    
    import matplotlib.pyplot as plt
    import numpy as np
    from PIL import Image
    from matplotlib.mlab import window_none
    
    
    def img2wav(img_path, wav_path, fft_size=1024):
        """ 把图片写到音频频域 :param img_path: 输入图片路径 :param wav_path: 输出音频路径 :param fft_size: 图片每列代表的音频长度，也是频域长度的两倍 """
    
        # 读取图片，转为灰度图
        img = Image.open(img_path).convert('L')
        # 缩放到高度 = fft_size / 2 （令负频率全为0）
        img = img.resize((img.width * fft_size // 2 // img.height, fft_size // 2),
                         Image.BICUBIC)
    
        # 转为numpy数组
        img = np.array(img, 'float')
        # 变换到-100~0分贝
        img = img * (100 / 255) - 100
        # 单位从分贝转成1，此时取值为0~1
        # amp_dB = 20 * ln(amp) / ln(10)
        # amp = exp(amp_dB / 20 * ln(10))
        img = np.exp(img * (np.log(10) / 20))
        # 翻转（索引小的频率小）然后转置（要迭代列）
        img = img[::-1].T
    
        # 防溢出，每列总振幅不能超过1
        max_sum = max(col.sum() for col in img)
        if max_sum > 1:
            img /= max_sum
    
        with wave.open(wav_path, 'wb') as f:
            # (nchannels, sampwidth, framerate, nframes, comptype, compname)
            f.setparams((1, 2, 44100, len(img) * fft_size, 'NONE', ''))
            for col in img:
                # 傅里叶反变换
                data = np.fft.ifft(col * fft_size, fft_size).real * 32767
    
                # 最后一次防溢出，限制范围在-32768~32767
                for index in np.where(data < -32768):
                    data[index] = -32768
                for index in np.where(data > 32767):
                    data[index] = 32767
    
                # 写到wav文件
                data = data.astype('short')
                f.writeframesraw(data)
    
    
    def draw_spectrum(wav_path, fft_size=1024):
        """ 画音频频谱图 :param wav_path: 输入音频路径 :param fft_size: 傅里叶变换用的长度 """
    
        # 读wav
        with wave.open(wav_path, 'rb') as f:
            n_samples = f.getnframes()
            data = f.readframes(n_samples)
            n_channels = f.getnchannels()
            sample_rate = f.getframerate()
    
        # 转为numpy数组
        data = np.fromstring(data, 'short')
        # 取第一个声道
        data.shape = (n_samples, n_channels)
        data = data.T[0]
    
        # 画频谱，无加窗和重叠
        plt.specgram(data / 32767, fft_size, sample_rate, window=window_none,
                     noverlap=0, scale='dB')
        plt.show()
    
    
    if __name__ == '__main__':
        img2wav('test.jpg', 'test.wav')
        draw_spectrum('test.wav')

测试图片：  
![lena][]  
![9][]  
![python][]

自己画的频谱效果  
![lena][lena 1]  
![9][9 1]  
![python][python 1]

AU的频谱效果，因为AU有加窗和重叠所以和自己画的有点不一样  
![lena][lena 2]  
![9][9 2]  
![python][python 2]

[GitHub]: https://github.com/xfgryujk/img2wav
[lena]: /images/20220603/9b6d453a80f54d70b37518fbb08351ef.png
[9]: /images/20220603/61994db888d94c0189550a77f05c033c.png
[python]: /images/20220603/07c77c851a4548f2bfc4c4e10f99542b.png
[lena 1]: /images/20220603/3a9156b897474d22b9678434089a2269.png
[9 1]: /images/20220603/b097f905a1e942059d6dbf15e3f83a85.png
[python 1]: /images/20220603/928cbc3c139a489192e7905e10f9eaf2.png
[lena 2]: /images/20220603/778f2e1f844849558eb50df079895329.png
[9 2]: /images/20220603/57f2791bc16a4afc86cc6144ec912007.png
[python 2]: /images/20220603/853419af202749aca9354acdd7106d2f.png