[kriboe90]

What changed

Step 3 analysed the channels of an image one after another, and each channel opened and decoded the image again. An RGB analysis (--rgb) therefore decoded every RAW file three times. Nothing the decoding yields depends on the channel: the black level and the read noise are taken per Bayer color, and only the channel mask is built per channel.

• ledsa/core/image_reading.py: the new ImageChannels reads the file once (decoding, black level and scaling, read noise, orientation). channel_data_mask_and_noise(channel) builds the array, mask and noise of one channel at a time from it, so that besides the decoded image only the channel in work is held in memory. read_channel_data_mask_and_noise_from_img uses it as well, so the reads for one and for several channels cannot drift apart.
• ledsa/data_extraction/DataExtractor.py: process_img_file reads the image before the channel loop and passes it to every channel.
• ledsa/data_extraction/step_3_functions.py: generate_analysis_data takes the image as the optional img_channels; without it, it reads the image itself as before.
• ledsa/tests/UnitTests/test_image_channels.py: 10 tests. For RAW (upright and sideways) and PNG, the single read gives exactly what a read per channel gives. Step 3 opens the file once for three channels. A grayscale or missing image is still reported as a corrupted image. The step 3 test fails against the old DataExtractor.

Why

On a 24 MP CR2 of demo_2, reading the three channels takes 0.42 s instead of 1.22 s per image, with bit-identical arrays, masks and noise for R, G and B. The share of this in the whole of step 3 was not measured: with -s3 the LED fitting probably dominates, with -s3_fast the read is a larger part.

For the reviewer

• An analysis of a single channel (the default, red) is unchanged.
• An image that cannot be read is now skipped for all channels at once, before any result file is written. Before, a failure in a later channel could leave the result files of the earlier ones, and --restart only checks channel 0.
• Peak memory per worker stays as it was: the per-channel arrays are built one at a time, never all three together.
• The single-channel reads in step 1 and in postprocessing/simulation.py are not touched.

Tests

• Unit tests: 378 passed
• Robot Framework acceptance tests: 32 passed

🤖 Generated with Claude Code