Brightest image pixel with neural network

Question

I’m trying to train a neural network to find the brightest pixel for each image in a dataset. I’ve started with a simple rgb images dataset with 3 random grey values:

def generate():
    a = random.randrange(0, 255)
    b = random.randrange(0, 255)
    c = random.randrange(0, 255)
    return a, b, c

def colors (a, b, c):
    for m in range(height):
           for n in range(30):
                   img.putpixel( (n,m), (a, a, a))        
    for m in range(height):
           for n in range(30,60):
                   img.putpixel( (n,m), (b, b, b))        
    for m in range(height):
           for n in range(60, 90):
                   img.putpixel( (n,m), (c, c, c))
    return img

for i in range (1000):   
    a, b, c = generate()
    width = 90
    height = 90
    img = Image.new('RGB', [width,height])
    img = colors(a,b,c)
    img.save('img/%d.bmp'%(i))

Then I create the array with the images and the label list with the highest pixel value:

import cv2
import glob
import numpy as np
import matplotlib as plt
import matplotlib.pyplot as plt

X = []
y = []

files = glob.glob ('img/*bmp')

for myFile in files:
    image = cv2.imread (myFile)
    biggest = np.amax(image)
    y.append (biggest)    
    X.append (image)

X = np.array(X)
X = X.astype('float32') / 255

print(X.shape)

I set the train/test sizes and transform the label list into an array:

from sklearn.model_selection import train_test_split    

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.33)
y_train = np.asarray(y_train)
y_test = np.asarray(y_test)

And here comes trouble: I can’t even understand what kind of modeling problem is, because it looks different from classification problems. Here I have a different target value for each image and if I try to adapt some models I’ve already used (adjusting the values) I got very high loss and very low accuracy. Here a model:

from tensorflow.keras import datasets, layers, models

model = models.Sequential()
model.add(layers.Conv2D(32, (11, 11), activation='relu', input_shape=(90, 90, 3)))
model.add(layers.MaxPooling2D(5,5))
model.add(layers.Conv2D(64, (9, 9), activation='relu',))
model.add(layers.MaxPooling2D(4,4))

model.add(layers.Flatten())
model.add(layers.Dense(64, activation='relu'))
model.add(layers.Dense(32, activation='relu'))
model.add(layers.Dense(255, activation = 'softmax'))


model.compile(optimizer='adam',
              
              loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
              metrics=['accuracy'])

model.summary()

model.fit(X_train, y_train,
                epochs=40,
                shuffle=True,
                validation_data=(X_test, y_test))

I also think I need to customize the loss function and the accuracy, because it’s not only a true/false problem: an output value close to the target can’t be considered as “false”. Now I’m not totally sure if the biggest problem is the training model or the loss function.

Brightest image pixel with neural network

0 Answers