Add a threshold for maximum amount of padding

This commit is contained in:
Arnaud_Cayrol 2026-02-15 16:59:47 +01:00
parent 1cdb1d7eb0
commit facc140619
7 changed files with 296 additions and 4 deletions

View file

@ -45,6 +45,10 @@
enabled: config.processing.face_resolution.enabled,
min_size: Number(config.processing.face_resolution.min_size),
},
crop: {
enabled: config.processing.crop.enabled,
max_padding_percent: Number(config.processing.crop.max_padding_percent),
},
brightness: {
enabled: config.processing.brightness.enabled,
min_brightness: Number(config.processing.brightness.min_brightness),
@ -203,6 +207,38 @@
{/if}
</div>
<!-- Face Too Close to Edge Section -->
<div class="setting-section">
<div class="section-header">
<span class="section-title">Face Too Close to Edge</span>
<input
type="checkbox"
bind:checked={config.processing.crop.enabled}
/>
</div>
<div class="section-hint">Faces too close to the edge of the photo are often distorted and require too much padding</div>
{#if config.processing.crop.enabled}
<div class="setting-row sub-setting">
<label for="max-padding">
<span class="setting-label">Max Padding</span>
<span class="setting-hint">Maximum allowed padding from edge replication</span>
</label>
<div class="setting-control">
<input
id="max-padding"
type="range"
bind:value={config.processing.crop.max_padding_percent}
min="5"
max="80"
step="5"
/>
<span class="value">{config.processing.crop.max_padding_percent}%</span>
</div>
</div>
{/if}
</div>
<!-- Brightness Section -->
<div class="setting-section">
<div class="section-header">
@ -793,6 +829,12 @@
color: #e0e0e0;
}
.section-hint {
font-size: 0.75rem;
color: #666;
margin-bottom: 0.5rem;
}
.section-header input[type='checkbox'] {
width: 1.25rem;
height: 1.25rem;

View file

@ -201,6 +201,41 @@ impl BlurConfig {
}
}
/// Crop padding filter configuration.
///
/// Skips images where too much of the crop region falls outside the original
/// image bounds (filled by edge-pixel replication).
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct CropConfig {
/// Whether crop padding filtering is enabled.
pub enabled: bool,
/// Maximum allowed padding percentage (0.0-100.0).
/// Images where the replicated-fill area exceeds this will be skipped.
pub max_padding_percent: f32,
}
impl Default for CropConfig {
fn default() -> Self {
Self {
enabled: false,
max_padding_percent: 30.0,
}
}
}
impl CropConfig {
/// Validate the configuration values.
pub fn validate(&self) -> Result<()> {
if self.enabled && !(0.0..=100.0).contains(&self.max_padding_percent) {
return Err(Error::Config(
"Crop max_padding_percent must be between 0.0 and 100.0".to_string(),
));
}
Ok(())
}
}
/// Output/resize configuration.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct OutputConfig {
@ -508,6 +543,10 @@ pub struct ProcessingConfig {
#[serde(default)]
pub face_resolution: FaceResolutionConfig,
/// Crop padding filter settings.
#[serde(default)]
pub crop: CropConfig,
/// Blur detection settings.
#[serde(default)]
pub blur: BlurConfig,
@ -547,6 +586,7 @@ impl Default for ProcessingConfig {
max_workers: 1,
use_preview: true,
face_resolution: FaceResolutionConfig::default(),
crop: CropConfig::default(),
blur: BlurConfig::default(),
brightness: BrightnessConfig::default(),
head_pose: HeadPoseConfig::default(),
@ -568,6 +608,7 @@ impl ProcessingConfig {
));
}
self.face_resolution.validate()?;
self.crop.validate()?;
self.blur.validate()?;
self.brightness.validate()?;
self.head_pose.validate()?;

View file

@ -9,6 +9,25 @@ use image::imageops::FilterType;
use image::{DynamicImage, GenericImageView, RgbImage};
use std::borrow::Cow;
/// Per-edge padding fractions (0.0-1.0) relative to crop size.
#[derive(Debug, Clone, Copy)]
pub struct PaddingEdges {
pub top: f32,
pub bottom: f32,
pub left: f32,
pub right: f32,
}
impl PaddingEdges {
/// Total fraction of crop area that is padding.
pub fn total_fraction(&self) -> f32 {
// Area = 1 - (1-left-right)*(1-top-bottom)
let inner_w = (1.0 - self.left - self.right).max(0.0);
let inner_h = (1.0 - self.top - self.bottom).max(0.0);
1.0 - inner_w * inner_h
}
}
/// Result of cropping a face from an image.
pub struct CropResult {
/// The cropped image at full resolution.
@ -17,6 +36,10 @@ pub struct CropResult {
pub resized: DynamicImage,
/// The face bounding box in crop coordinates.
pub face_rect: BoundingBox,
/// Fraction of the crop area that falls outside the original image (0.0-1.0).
pub padding_fraction: f32,
/// Per-edge padding fractions.
pub padding_edges: PaddingEdges,
}
/// Crop and resize the face from an image using bounding box.
@ -70,6 +93,25 @@ pub fn crop_face_with_intermediate(
let ideal_x1 = center_x as i32 - crop_size as i32 / 2;
let ideal_y1 = center_y as i32 - crop_size as i32 / 2;
// Calculate what fraction of the crop area falls outside the image bounds.
let overlap_x1 = (ideal_x1).max(0) as u64;
let overlap_y1 = (ideal_y1).max(0) as u64;
let overlap_x2 = ((ideal_x1 + crop_size as i32) as u64).min(img_width as u64);
let overlap_y2 = ((ideal_y1 + crop_size as i32) as u64).min(img_height as u64);
let overlap_area =
overlap_x2.saturating_sub(overlap_x1) * overlap_y2.saturating_sub(overlap_y1);
let total_area = crop_size as u64 * crop_size as u64;
let padding_fraction = 1.0 - (overlap_area as f32 / total_area as f32);
// Per-edge padding as fraction of crop size
let cs = crop_size as f32;
let padding_edges = PaddingEdges {
left: (-ideal_x1).max(0) as f32 / cs,
top: (-ideal_y1).max(0) as f32 / cs,
right: ((ideal_x1 + crop_size as i32) - img_width as i32).max(0) as f32 / cs,
bottom: ((ideal_y1 + crop_size as i32) - img_height as i32).max(0) as f32 / cs,
};
// Crop with replicate-fill: always keeps the face centered by extending
// edge pixels at image borders instead of shifting the crop window.
let cropped = crop_with_replicate_fill(img, ideal_x1, ideal_y1, crop_size);
@ -89,6 +131,8 @@ pub fn crop_face_with_intermediate(
cropped,
resized,
face_rect,
padding_fraction,
padding_edges,
})
}

View file

@ -20,7 +20,7 @@ pub mod steps;
mod traits;
mod types;
pub use crop_utils::{crop_face_with_intermediate, CropResult};
pub use crop_utils::{crop_face_with_intermediate, CropResult, PaddingEdges};
pub use debug_utils::{draw_simple_text, face_rect_pixels};
pub use orientation::load_image_with_orientation;
pub use traits::*;

View file

@ -5,10 +5,12 @@
use crate::config::Config;
use crate::pipeline::crop_face_with_intermediate;
use crate::pipeline::debug_utils::draw_simple_text;
use crate::pipeline::{
computed_keys, BoundingBox, ComputedValue, PipelineContext, ProcessingStep, StepOutcome,
};
use async_trait::async_trait;
use image::{DynamicImage, Rgb};
/// Crops the face region from the full image and resizes it.
///
@ -39,6 +41,30 @@ impl ProcessingStep for CropAndResizeStep {
// Crop returns CropResult with cropped images and face rectangle in crop coordinates
match crop_face_with_intermediate(image, &ctx.face_data, output_size, eye_distance) {
Ok(crop_result) => {
// Store padding info for debug visualization
ctx.set_computed(
computed_keys::PADDING_EDGES,
ComputedValue::PaddingEdges(crop_result.padding_edges),
);
// Check if too much of the crop falls outside the image
let crop_config = &config.processing.crop;
if crop_config.enabled {
let padding_pct = crop_result.padding_fraction * 100.0;
if padding_pct > crop_config.max_padding_percent {
// Set image so debug_visualize can use it
ctx.image = Some(crop_result.resized);
return StepOutcome::Skip {
ctx,
reason: "excessive_padding".to_string(),
detail: Some(format!(
"padding {:.1}% exceeds max {:.1}%",
padding_pct, crop_config.max_padding_percent
)),
};
}
}
// Use the pre-resized image from the crop function
let cropped_size = crop_result.cropped.width();
ctx.image = Some(crop_result.resized);
@ -67,6 +93,117 @@ impl ProcessingStep for CropAndResizeStep {
},
}
}
fn debug_visualize(&self, ctx: &PipelineContext, config: &Config) -> Option<DynamicImage> {
let edges = ctx
.get_computed(computed_keys::PADDING_EDGES)
.and_then(|v| v.as_padding_edges())?;
let image = ctx.image.as_ref()?;
let rgb = image.to_rgb8();
let (width, height) = (rgb.width(), rgb.height());
let mut debug_img = rgb.clone();
// Tint padded regions with a semi-transparent red overlay
let tint = |pixel: &Rgb<u8>| -> Rgb<u8> {
Rgb([
(pixel[0] as u16 / 2 + 127).min(255) as u8,
pixel[1] / 2,
pixel[2] / 2,
])
};
let left_px = (edges.left * width as f32).round() as u32;
let right_px = (edges.right * width as f32).round() as u32;
let top_px = (edges.top * height as f32).round() as u32;
let bottom_px = (edges.bottom * height as f32).round() as u32;
// Tint left edge
for y in 0..height {
for x in 0..left_px.min(width) {
debug_img.put_pixel(x, y, tint(debug_img.get_pixel(x, y)));
}
}
// Tint right edge
for y in 0..height {
for x in width.saturating_sub(right_px)..width {
debug_img.put_pixel(x, y, tint(debug_img.get_pixel(x, y)));
}
}
// Tint top edge (only the non-corner part to avoid double-tinting)
for y in 0..top_px.min(height) {
for x in left_px.min(width)..width.saturating_sub(right_px) {
debug_img.put_pixel(x, y, tint(debug_img.get_pixel(x, y)));
}
}
// Tint bottom edge (only the non-corner part)
for y in height.saturating_sub(bottom_px)..height {
for x in left_px.min(width)..width.saturating_sub(right_px) {
debug_img.put_pixel(x, y, tint(debug_img.get_pixel(x, y)));
}
}
// Draw padding percentage bar at the bottom
let total_pct = edges.total_fraction() * 100.0;
let max_pct = config.processing.crop.max_padding_percent;
let bar_height = 20u32;
let bar_y = height.saturating_sub(bar_height);
let bar_width = (width as f32 * 0.8) as u32;
let bar_x = (width - bar_width) / 2;
// Background
for y in bar_y..height {
for x in 0..width {
debug_img.put_pixel(x, y, Rgb([40, 40, 40]));
}
}
// Bar outline
let outline_y = bar_y + 4;
let outline_height = bar_height - 8;
for x in bar_x..bar_x + bar_width {
debug_img.put_pixel(x, outline_y, Rgb([200, 200, 200]));
debug_img.put_pixel(x, outline_y + outline_height - 1, Rgb([200, 200, 200]));
}
for y in outline_y..outline_y + outline_height {
debug_img.put_pixel(bar_x, y, Rgb([200, 200, 200]));
debug_img.put_pixel(bar_x + bar_width - 1, y, Rgb([200, 200, 200]));
}
// Fill bar (scale: 0-50% maps to full bar)
let max_scale = 50.0_f32;
let normalized = (total_pct / max_scale).clamp(0.0, 1.0);
let fill_width = ((bar_width - 4) as f32 * normalized) as u32;
let fill_color = if total_pct > max_pct {
Rgb([255, 80, 80]) // Red - exceeds threshold
} else if total_pct > max_pct * 0.7 {
Rgb([255, 200, 80]) // Yellow - approaching threshold
} else {
Rgb([80, 255, 80]) // Green - well within threshold
};
for y in (outline_y + 2)..(outline_y + outline_height - 2) {
for x in (bar_x + 2)..(bar_x + 2 + fill_width) {
if x < width {
debug_img.put_pixel(x, y, fill_color);
}
}
}
// Draw threshold marker on the bar
let threshold_x = bar_x + 2 + ((bar_width - 4) as f32 * (max_pct / max_scale).clamp(0.0, 1.0)) as u32;
if threshold_x < bar_x + bar_width {
for y in outline_y..(outline_y + outline_height) {
debug_img.put_pixel(threshold_x, y, Rgb([255, 255, 255]));
}
}
// Text label
let text = format!("{:.1}%", total_pct);
draw_simple_text(&mut debug_img, 5, bar_y + 6, &text, Rgb([255, 255, 255]));
Some(DynamicImage::ImageRgb8(debug_img))
}
}
#[cfg(test)]

View file

@ -29,6 +29,8 @@ pub mod computed_keys {
pub const HEAD_POSE: &str = "head_pose";
/// Face bounding box in current image coordinates computed by CropAndResizeStep.
pub const FACE_RECT: &str = "face_rect";
/// Per-edge padding fractions computed by CropAndResizeStep.
pub const PADDING_EDGES: &str = "padding_edges";
}
/// Outcome of a pipeline step execution.
@ -63,6 +65,8 @@ pub enum ComputedValue {
Landmarks(Box<Landmarks>),
/// Face bounding box in current image coordinates.
FaceRect(BoundingBox),
/// Per-edge padding fractions from crop step.
PaddingEdges(super::PaddingEdges),
}
impl ComputedValue {
@ -121,6 +125,14 @@ impl ComputedValue {
_ => None,
}
}
/// Get as PaddingEdges if this is a PaddingEdges variant.
pub fn as_padding_edges(&self) -> Option<&super::PaddingEdges> {
match self {
ComputedValue::PaddingEdges(v) => Some(v),
_ => None,
}
}
}
/// A debug image with metadata about whether the step passed or failed.

View file

@ -1,9 +1,9 @@
//! Configuration endpoints.
use crate::config::{
AlignmentConfig, BlurConfig, BrightnessConfig, EyeFilterConfig, FaceResolutionConfig,
HeadPoseConfig, OutputConfig, ProcessingConfig, TimeIntervalConfig, TimestampConfig,
VideoConfig, CONFIG_PATH,
AlignmentConfig, BlurConfig, BrightnessConfig, CropConfig, EyeFilterConfig,
FaceResolutionConfig, HeadPoseConfig, OutputConfig, ProcessingConfig, TimeIntervalConfig,
TimestampConfig, VideoConfig, CONFIG_PATH,
};
use crate::web::state::AppState;
use axum::{extract::State, http::StatusCode, response::Json};
@ -48,6 +48,7 @@ pub struct ConfigUpdateRequest {
pub struct ProcessingConfigUpdate {
pub max_workers: Option<usize>,
pub face_resolution: Option<FaceResolutionConfig>,
pub crop: Option<CropConfig>,
pub blur: Option<BlurConfig>,
pub brightness: Option<BrightnessConfig>,
pub head_pose: Option<HeadPoseConfig>,
@ -108,6 +109,18 @@ fn validate_processing_config(proc: &ProcessingConfigUpdate) -> Result<(), Valid
}
}
if let Some(ref cr) = proc.crop {
if cr.enabled && !(0.0..=100.0).contains(&cr.max_padding_percent) {
return Err(ValidationError::new(
"processing.crop.max_padding_percent",
format!(
"must be between 0.0 and 100.0, got {}",
cr.max_padding_percent
),
));
}
}
if let Some(ref bl) = proc.blur {
if bl.enabled && bl.min_sharpness < 0.0 {
return Err(ValidationError::new(
@ -268,6 +281,9 @@ pub async fn update_config(
if let Some(v) = proc.face_resolution {
config.processing.face_resolution = v;
}
if let Some(v) = proc.crop {
config.processing.crop = v;
}
if let Some(v) = proc.blur {
config.processing.blur = v;
}