Target recognition

Two common ways of getting from sidescan data to machine-learning target recognition, step by step. Each step links to an example that does it; sidescanpy covers the sonar side (reading, correcting, cutting images, mapping image boxes back to the data), and established tools do the learning and the annotation.

From raw data to a two-stage recognizer

Most sidescan target recognition runs in two stages: a detector finds anything on the seabed in whole swath images and may raise false alarms, then a classifier looks at a small chip around each detection and decides what it is, or rejects it.

Step	What happens	Example
1. Convert and correct	raw JSF, SDF or XTF to the data model, then TVG, ground range and positions, one call per correction	Getting started, Correct a line
2. Cut images to annotate	dual-swath tiles with square pixels in metres, plus an index of where each tile sits in the data	Swath tiles for annotation
3. Annotate	draw boxes in FiftyOne, CVAT or Label Studio and export COCO; read the boxes back as targets with ping, side and position	Swath tiles for annotation, Review and correct labels
4. Train a detector	one class, “object”, on whole images	Labelled images, Train a detector
5. Train a classifier	chips around each labelled object, plus clutter and background images	Train a classifier, Training dataset with backgrounds
6. Find candidates	run both stages on new tiles and turn the detections into targets with positions	Run the models on a survey
7. Review	confirm, reject and correct the candidates; the corrections are the next round’s training data	Review and correct labels
8. Test on more data	score on lines or surveys that were never used for training, in image space or in metres	Test both stages, Run the models on a survey

The step-by-step pages walk through steps 4 to 8 in turn. The detector and classifier examples train on a public labelled image set (Santos et al. 2024, mine-like objects seen by a Marine Sonic sonar), because no openly licensed raw sidescan recording with labelled targets was available. With your own data, steps 2 and 3 produce the same kind of images and labels.

From a target catalog to image chips

When a survey already has picked targets, from acquisition software (EdgeTech Discover and Klein catalogs are read directly) or from an analyst, the chips come straight from the catalog.

Step	What happens	Example or function
1. Read the catalog	into Target records; keep it as JSON or CSV	io.read_target_catalog, targets.read_targets_json (Targets and datasets)
2. Place targets on pings	by recorded file and ping, by position, or by ping number	targets.correlate (Targets and datasets)
3. Render chips	at any correction level, pixel size and window style, with COCO and YOLO labels	Target chips at several corrections and resolutions
4. Add backgrounds and split	images with no target, and a train/validation split that keeps overlapping windows together	Training dataset with backgrounds

Building blocks

Function	Does
export.ml.write_swath_tiles	a corrected channel as PNG swath tiles plus tiles.json
export.ml.read_coco_boxes	boxes from a COCO file (annotations or detector results), per image
export.ml.boxes_to_targets	boxes on a tile to targets: file, ping, side, sample, position, size in metres, score
export.ml.build_target_dataset	chips and background images with COCO and YOLO labels, at any correction level, resolution and style
targets.match_targets	candidates against a catalog within a distance: hits, false alarms, misses, precision, recall

Things that go wrong

• Splitting at random. Neighbouring swath images overlap and a target is often seen on several passes, so a random split puts near-copies in training and test and flatters the scores. Split by line, survey or date.

• Backgrounds that differ from the targets’ surroundings. If background images come from different ranges or look different after the display stretch, a classifier learns that instead of the objects. Look at them side by side before training.

• Another sonar. A model trained on one sonar, frequency or recorder setting rarely transfers to another without new labels. Correcting the data first (TVG, ground range, the same pixel size) narrows the gap but does not close it.