Squeeze and Excitation Rank Faster R-CNN for Ship Detection in SAR Images

if you want the project pls call @8125424511

Squeeze and Excitation Rank Faster R-CNN for Ship Detection in SAR Images

In this project author is describing concept to detect ships from sea images taken from satellites and this images are called as ‘Synthetic aperture radar (SAR)’. Ships can be detected from SAR images using Faster R-CNN (Region Based Convolution Neural Networks) Algorithm. RCNN algorithm will be trained with ship images from VGG ImageNet Network, while training RCNN extract features from images using its height, width and image colour channel. RCNN filter train images features map from multiple layers of convolution neural network. All object detection from image will be maintained in train vector with value 1 and other background features marked as 0. Whenever new SAR test image uploaded then RCNN algorithm will apply train vector on SAR test image to detect objects with ships features.

All existing algorithms required manual sea segmentation images from hands to detect ships, if segmentation is not accurate or blur is capture in test images then ships will be detected falsely or no detection will happen. To overcome from this issue author is using deep learning FASTER RCNN algorithm with Squeeze and Excitation concept.

Squeeze and Excitation concept is used to increase detection performance. RCNN extract features from train images and thenExcitation vector take relevant object detection features and ranked them and then only top Kvalues will be preserved with 1. Other values will be set to 0. Then,the redundant sub-feature maps are suppressed using squeeze technique. In simple terms Excitation will identify top K features and accept them and squeeze technique remove all irrelevant features. After applying Squeeze and Excitation RCNN will have train model with important relevant features. When we give new test image then that model apply on test image to detect ships.

RCNN Working Procedure

After extracting shared feature maps from train images with a CNN, the firststageRegion Proposal Network (RPN) takes shared feature maps as input and generates aset of rectangular candidate object locations named anchors,each with an objectness score. The size of each anchordepends on hyperparameters named scales and aspect ratios.A small network is slid over the shared feature maps and mappingeach sliding window to lower dimensional feature. Aftertwo sibling fully connected layers, the 2A-dimensional vectorbox-classification and 4A-dimensional vector box-regressionare generated for A anchors of each point on the feature map.

Modules Information

This project consists of following modules

1)    Generate Faster RCNN Model: In this module a train RCNN model will be generated using Squeeze and Excitation. Input data to this module is given from ‘VGGImageNet.h5py’ file. RCNN extract all features from this file and build a train model, while building model it will read all data from file and then using Squeeze and Excitation technique accept top K features and squeeze irrelevant features. RCCN internally uses CNN pooling technique to build mode. Below code describe model generation for train images and use five convolution layers.

In below code input_shape refers to image size as 80 = height, 80 = width and 3 means given image is a colour image

model.add(Conv2D(32, (3, 3), padding='same', input_shape=(80, 80, 3), activation='relu'))

model.add(MaxPooling2D(pool_size=(2, 2))) #40x40

model.add(Dropout(0.25))  //layer 1 object

model.add(Conv2D(32, (3, 3), padding='same', activation='relu'))

model.add(MaxPooling2D(pool_size=(2, 2))) #20x20

model.add(Dropout(0.25)) //layer 2 object

model.add(Conv2D(32, (3, 3), padding='same', activation='relu'))

model.add(MaxPooling2D(pool_size=(2, 2))) #10x10

model.add(Dropout(0.25))

model.add(Conv2D(32, (10, 10), padding='same', activation='relu'))

model.add(MaxPooling2D(pool_size=(2, 2))) #5x5

model.add(Dropout(0.25))

model.add(Flatten())

model.add(Dense(512, activation='relu'))

model.add(Dropout(0.5))

In project code you can see complete program of training with Fast RCNN.

2)    Upload Test Image & Detect Ship: In this module we will upload test image and then application extract features from this test image and then apply RCNN train model on that test image to detect ships.

Author : ss infotech

Share this

Related Posts