AUTOMATIC FORENSIC SKETCH TO IMAGE GENERATOR USING GAN
ABSTARCT :
Image processing has been a crucial tool for refining the image or to boost the image. With the improvement of machine learning tools, the image processing task has been simplified to great range. Generating a semantic and photographic face image, from a sketch image or text description has always been an extremely important issue. Sketch images basically contain only simple profile information but not the detail of the face. Therefore, it is difficult to come up with facial characteristics accurately. In order to solve this problem, we propose an image translation network by exploiting attributes with the generated adversarial network (GAN). The generator network consists of a feature extracting network and down sampling–up sampling network. Both networks use skip-connection to reduce the number of layers without affecting network performance. The discriminator network is designed to inspect whether the generated faces contain the desired attributes or not.This way, the realistic photograph for any sketch can be obtained easily with exact detail and in less time.
EXISTING SYSTEM :
• This would give a distribution matching error that could be used to update the network via back propagation. This direct method is practically very complex to implement.
• A GAN has three primary components: a generator model for generating new data, a discriminator model for classifying whether generated data are real faces, or fake, and the adversarial network that pits them against each other.
• The generative part is responsible for taking N-dimensional uniform random variables (noise) as input and generating fake faces. The generator captures the probability P(X), where X is the input.
• The discriminative part is a simple classifier that evaluates and distinguished the generated faces from true celebrity faces. The discriminator captures the conditional probability P(Y|X), where X is the input and Y is the label.
DISADVANTAGE :
• Human-made sketches may be ambiguous and unclear.
• The primitive sketches of a target may be black and? white or monochrome.
• Features which are colour dependent will not be significant.
• Some features can be changed by culprit/target. E.g.? hair colour, hairstyle, etc.
• Face recognition by humans takes more time if the? target face has to find a match without a database.
PROPOSED SYSTEM :
• Similar to face hallucination reconstruction, we propose anew network based on GAN.
• In the feature extraction stage,we extract profile information and high-level semantic information from sketch images and attribute vectors.
• A combination of GAN and Skip connection layers are used in all the generator.
• Each output of the convolution layer is passed to the back convolution layer and simultaneously concatenated to the next concatenation layer.
ADVANTAGE :
• Easy recognition
• Accommodates ambiguous input faces
• Efficient with increased use
• Fast and accurate results
• Scope for broader applications
|