✨ feat(predict): add Error Level Analysis to image prediction

【Predict Image Function】
- integrate an extra transform using ELA (Error Level Analysis)

【User Interface】
- display ELA transformed image

【Settings】
- updated the visibility of settings slider

app.py

@@ -7,7 +7,7 @@ from PIL import Image
 import numpy as np
 import io
 import logging
-from utils.utils import softmax, augment_image, convert_pil_to_bytes
+from utils.utils import softmax, augment_image, convert_pil_to_bytes, ELA
 
 
 # Configure logging
@@ -219,8 +219,9 @@ def predict_image_with_html(img, confidence_threshold, augment_methods, rotate_d
     else:
         img_pil = img
     img_pil, results = predict_image(img_pil, confidence_threshold)
+    ela_img = ELA(img_pil)  # Apply ELA to the image
     html_content = generate_results_html(results)
-    return img_pil, html_content
+    return img_pil, ela_img, html_content
 
 with gr.Blocks() as iface:
     with gr.Tab("AI Image Detection"):
@@ -231,20 +232,22 @@ with gr.Blocks() as iface:
                 image_input = gr.Image(label="Upload Image to Analyze", sources=['upload'], type='pil')
                 with gr.Accordion("Settings", open=False, elem_id="settings_accordion"):
                     augment_checkboxgroup = gr.CheckboxGroup(["rotate", "add_noise", "sharpen"], label="Augmentation Methods")
-                    rotate_slider = gr.Slider(0, 360, value=0, step=1, label="Rotate Degrees", visible=False)
-                    noise_slider = gr.Slider(0, 100, value=0, step=1, label="Noise Level", visible=False)
-                    sharpen_slider = gr.Slider(0, 100, value=1, step=1, label="Sharpen Strength", visible=False)
-                    confidence_slider = gr.Slider(0.0, 1.0, value=0.5, step=0.01, label="Confidence Threshold")
+                    rotate_slider = gr.Slider(0, 45, value=2, step=1, label="Rotate Degrees", visible=False)
+                    noise_slider = gr.Slider(0, 50, value=4, step=1, label="Noise Level", visible=False)
+                    sharpen_slider = gr.Slider(0, 50, value=11, step=1, label="Sharpen Strength", visible=False)
+                    confidence_slider = gr.Slider(0.0, 1.0, value=0.75, step=0.05, label="Confidence Threshold")
                 inputs = [image_input, confidence_slider, augment_checkboxgroup, rotate_slider, noise_slider, sharpen_slider]
                 predict_button = gr.Button("Predict")
                 image_output = gr.Image(label="Processed Image", visible=True)
+                ela_image_output = gr.Image(label="ELA Processed Image", visible=True)
+
 
             with gr.Column(scale=2):
                 with gr.Accordion("Project OpenSight - Model Evaluations & Playground", open=False, elem_id="project_accordion"):
                     gr.Markdown("## OpenSight is a SOTA gen. image detection model, in pre-release prep.\n\nThis HF Space is a temporary home for us and the public to evaluate the shortcomings of current open source models.\n\n<-- Feel free to play around by starting with an image as we prepare our formal announcement.")
                 # Custom HTML component to display results in 5 columns
                 results_html = gr.HTML(label="Model Predictions")
-                outputs = [image_output, results_html]
+                outputs = [image_output, ela_image_output, results_html]
         
         # Show/hide rotate slider based on selected augmentation method
         augment_checkboxgroup.change(lambda methods: gr.update(visible="rotate" in methods), inputs=[augment_checkboxgroup], outputs=[rotate_slider])

utils/utils.py

@@ -1,6 +1,6 @@
 import numpy as np
 import io
-from PIL import Image, ImageFilter
+from PIL import Image, ImageFilter, ImageChops
 from torchvision import transforms
 
 def softmax(vector):
@@ -22,4 +22,21 @@ def convert_pil_to_bytes(image, format='JPEG'):
     img_byte_arr = io.BytesIO()
     image.save(img_byte_arr, format=format)
     img_byte_arr = img_byte_arr.getvalue()
-    return img_byte_arr
+    return img_byte_arr
+
+def ELA(img_pil, scale=77, alpha=0.66):
+    # Error Level Analysis for basic image forensics
+    original = img_pil.copy()          # open up the input image
+    temp_path = 'temp.jpg'             # temporary image name to save the ELA to
+    original.save(temp_path, quality=95)     # re-save the image with a quality of 95%
+    temporary = Image.open(temp_path)        # open up the re-saved image
+
+    diff = ImageChops.difference(original, temporary)   # load in the images to look at pixel by pixel differences
+    d = diff.load()                     # load the image into a variable
+    WIDTH, HEIGHT = diff.size           # set the size into a tuple
+    for x in range(WIDTH):                                  # row by row
+        for y in range(HEIGHT):                             # column by column
+            d[x, y] = tuple(k * scale for k in d[x, y])     # set the pixels to their x,y & color based on error
+
+    new_img = ImageChops.blend(temporary, diff, alpha)      # blend the original w/ the ELA @ a set alpha/transparency
+    return new_img