Scribd
Upload
75 views2 pages

EfficientNet Architectures Overview

EfficientNet is a family of deep learning architectures optimized for image processing, utilizing a compound scaling method to enhance performance while minimizing parameters. EfficientNet-B3 is particularly suitable for image retrieval tasks, while models B0 to B2 are too small for high-quality retrieval, and B4 to B7 are excessive for the CIFAR-100 dataset. HNSW is an efficient graph-based algorithm for approximate nearest neighbors search, building a multi-layer graph to connect similar images for quick retrieval.

Uploaded by

M. Roshini
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
75 views2 pages

EfficientNet Architectures Overview

EfficientNet is a family of deep learning architectures optimized for image processing, utilizing a compound scaling method to enhance performance while minimizing parameters. EfficientNet-B3 is particularly suitable for image retrieval tasks, while models B0 to B2 are too small for high-quality retrieval, and B4 to B7 are excessive for the CIFAR-100 dataset. HNSW is an efficient graph-based algorithm for approximate nearest neighbors search, building a multi-layer graph to connect similar images for quick retrieval.

Uploaded by

M. Roshini
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd

EfficientNet

EfficientNet Models are a family of deep learning architectures designed for efficient and
accurate image processing. It uses a compound scaling method to balance depth,
width, and resolution for better performance with fewer parameters.

Compound scaling method:


The Compound Scaling Method is a technique used in EfficientNet to systematically
scale depth, width, and resolution of a neural network together to improve performance
efficiently.

Depth(more layers):Learns complex features


Width(more channels):captures more details
Resolution (Bigger Images):Better feature extraction

EfficientNet Variants:

Model Parameters

EfficientNet-B0 5.3 M

EfficientNet-B1 7.8 M

EfficientNet-B2 9.2 M

EfficientNet-B3 12 M

EfficientNet-B4 19 M

EfficientNet-B5 30 M

EfficientNet-B6 43 M

EfficientNet-B7 66 M

●​ EfficientNet-B0 to B2 will be not ideal for image retrieval as they are too small
and may miss fine details in images, making them less useful for high-quality
retrieval.
EfficientNet-B3 is ideal for image retrieval task
●​ Extracts high-quality embeddings for image similarity search.
●​ CIFAR-100 has small 32×32 images, but we can resize them to 300×300 to
leverage EfficientNet's strength.(Handles Cifar 100 dataset well)
●​ Works well with ANN as it is ideal for fast retrieval

EfficientNet B4/B5/B6/B7 is not because they overkill for CIFAR-100 since images are
small and dataset size is only 50,000 training [Link] CIFAR-100 images are only
32×32 pixels they dont get additional useful images even after upscaling.

HNSW

HNSW (Hierarchical Navigable Small World) is a graph-based Approximate Nearest


Neighbors (ANN) search algorithm that efficiently finds similar items in large-scale
datasets.

●​ It uses graph structure to connect images based on similarity.

Working:
Build the Graph
●​ Each image (feature vector) is inserted into a multi-layer graph.
●​ Similar images are linked based on nearest neighbors.
●​ The top layers represent high-level grouping, while lower layers store
fine-grained details.

Search for Similar Images


●​ Start from top layers (global view) and move down to refine the search.
●​ Uses greedy search to find the closest match quickly

You might also like