Method and electronic device for generating a machine learning training dataset for shadow direction and removal

A client device for generating a training dataset for training a machine learning (ML) model to perform image shadow detection, the client device comprising:

• an image capture device, for capturing a plurality of images of scenes;
• storage comprising a plurality of databases forming or being used to generate a training dataset for training the ML model;
• and
• at least one processor coupled to memory and arranged to:
• use a pre-trained shadow detection model to generate a training dataset by:
• obtaining a shadow mask for each image of the plurality of images;
• determining whether the obtained shadow mask satisfies criteria for adding the shadow mask to a database of shadow masks stored in the storage on the client device;
• and
• based on the criteria being satisfied by the obtained shadow mask for an image:
• adding the obtained shadow mask to the database of the shadow masks;
• detecting at least one shadow-free image region in the image;
• and
• extracting and adding the at least one shadow-free image region to a database of shadow-free patches stored in the storage on the client device.

A computer-implemented method, performed by a client device, for training a machine learning (ML) model using federated learning to perform image shadow detection, the method comprising:

• obtaining a pre-trained shadow detection model from a central server;
• generate, using the pre-trained shadow detection model, a training dataset by:
• obtaining a plurality of images of scenes;
• and
• obtaining a shadow mask for each image of the plurality of images;
• determining whether the obtained shadow mask satisfies criteria for adding the shadow mask to a database of shadow masks on the client device, the database of the shadow masks forming or being used to generate the training dataset;
• and
• based on the criteria being satisfied by the obtained shadow mask for an image:
• adding the obtained shadow mask to the database of the shadow masks;
• detecting at least one shadow-free image region in the image;
• and
• extracting and adding the at least one shadow-free image region to a database of shadow-free patches, the database of the shadow-free patches forming or being used to generate the training dataset;
• using the generated training dataset to train the ML model;
• sharing model weights of the trained ML model with the central server to update the ML model;
• and
• receiving the updated ML model from the central server.