Artificial intelligence for detecting waste accumulation

This archive includes automatic translations. Audio, video and text inside images remain in their original language.

Figure 1.- Accumulation of garbage in the streets.

Latam online. Second Edition. 2021

Currently, municipal waste worldwide could increase by 70% in the next 30 years, thus causing piles of garbage to accumulate around the world (World Bank, 2019). This is why, if urgent measures are not taken, a future is expected where living with garbage is normal for society.

Due to the consequences of uncontrolled and unplanned garbage accumulations, different strategies must be taken to cushion their occurrences in different parts of the world.

Problem description

The problem of the accumulation of garbage not only lies in the bad smell that is perceived, but also in the image of unhealthiness, disorder and even insecurity that the garbage scattered in the streets generates. In the streets you can find everything from cardboard and diapers to fruit peels, plastic and old sheets. All this at the foot of the rainwater channel, at whose bases it is also common to see all types of debris floating.

According to the Ministry of the Environment, in Ecuador the most critical is the Coast, where only 10% of the municipalities have a sanitary landfill, manual or mechanized; in the Sierra region they are 25%, and in the East, 24%. Therefore, the garbage figures within the country increase daily as seen in Figure 2.

Figure 2.- Garbage figures in Ecuador.

Figure 2.1 shows some of the consequences of the accumulation of garbage, such as: damage to public infrastructure, increase in diseases and pests, flooding in the streets, obstruction in sewers and among others.

Figure 2.1.- Consequences of garbage accumulation.

In the city of Guayaquil, penalties for disposing of non-hazardous solid waste outside the established frequency and schedule and accumulating garbage in flower beds and sidewalks; they range from $80 to $500 .

How is our project born?

Clean City was based on the following question : How to identify in a timely manner the accumulation of uncontrolled garbage in the city of Guayaquil to minimize the impact on society? For this reason we created an application that allows us to contribute to garbage collection. Involving citizens in improving the cleanliness of the city and getting some companies interested in our technological tool based on Artificial Intelligence.

Project members

Figure 3.- Purple team — SaturdayAI –EPIC 2021

Team experience

Manuel Ahumada "It was a practical learning experience: focused on doing to acquire the topics and concepts presented. That reinforces knowledge."

Patricia Andrade "My experience was very enriching due to all the new knowledge I learned. In addition, I really liked sharing ideas with colleagues from different fields to solve a practical problem."

Cesar Villarroel "It was a practical learning experience: focused on doing to achieve the acquisition of the topics and concepts presented. That reinforces knowledge."

Ingrid Leon "The experience obtained in this modality is good. Being able to share knowledge, challenges and risks makes us more professional. I am left with a phrase "There are two basic options: accept the conditions as they exist or accept the responsibility of modifying them", then we can always improve."

General objective

In the following figure 4, you can see the main idea of the project.

Develop an application based on Artificial Intelligence that allows identifying agglomerations of garbage on public roads in urban areas.

Figure 4.- Project value proposition.

Solution approach

It is projected that through the collaboration of citizens, awareness will be created where they are part of the cleanliness and care of the city. How are they going to do it? Easy, through the app “Clean City” , where they can take a photo to identify and report the accumulation of garbage in a certain sector. Additionally, the application records the date, time and location of the problem. This application will be based on Artificial Intelligence that will allow identifying agglomerations of garbage on public roads in urban areas and promoting an immediate action plan for health authorities.

Figure 5.- General Diagram of the solution.

Data set

The data set that was formed was based on downloading images individually from the Internet, videos transformed into frames. In addition, the Kaggle and Google Street view platform was released where a large number of datasets of garbage agglomeration images could be obtained.

Figure 6.- Means to obtain the data set

Model selection

For the selection of the models, the following considerations were obtained:

Easy implementation.
Model performance.
Calculation limitation.

These considerations are due to the resources available to achieve the general objective. The models to be developed are:

CCN Custom.
VGG-16.
homeV3.

Techniques implemented

The main technique used to compensate for the limited data set size due to the Google image search limitation was “data augmentation” performed by the “ImageDataGenerator” keras library.

Data augmentation

Data augmentation is the artificial generation of data through perturbations to the original data. This allows us to increase both the size and diversity of our training data set. In computer vision, this technique became a regularization standard, and also to improve performance and combat overfitting in CNNs.

Figure 7.- Example of data augmentation.

In the three selected models, the data augmentation technique was considered with the objective of normalizing or re-scaling the pixels in a range from 0 to 1. In addition to modifying the images of the train set with the objective that there is a distinction in each of the selected images in terms of width, length, amplitude, rotation and scaling. Below is the code where the training images are transformed with the selected parameters:

Figure 8.- Image training code.

Custom CNN

In the personalized CNN model, regulation techniques, callbacks and finally keras tuner were applied.

Keras Tuner

Keras Tuner is a very easy-to-use library that greatly simplifies the complexity process of applying hyper-parameter optimization to deep neural networks built using Keras, offering us a wide degree of versatility to optimize both the structure of our network and the configuration of the parameters of some of the algorithms involved in the training process.

Figure 9.- the value of the test and accuracy can be verified

While analyzing each trial, Keras Tuner verifies the best val accuracy value, inferring that it is the best configuration of the architecture, and proceeds to save the best configuration of the model architecture (number of layers) with which it tends to achieve the best metric.

Figure 10.- cnn model code

Custom cnn model architecture features regularizers, convolution, keras tuner

InceptionV3

Figure 11.- shows the pre-training

The inceptionv3 pre-trained model is defined, validating the size of the input images with which it is trained:

Figure 12.- use of the mixed7 architecture

The architecture is frozen in the “mixed7” layer and then altered by adding layers based on the prediction of our model, which is a binary classification by which the last layer has sigmoidal activation which offers us the probability of whether a case is positive or negative.

VGG-16

Figure 13.- Additional features with VGG16

The pre-trained vgg16 model is defined and validates the size of the input images with which it is trained

Figure 14.- increase in layers (binary classification)

We proceed to increase the layers based on the prediction of our model, which is a binary classification, so the last layer has sigmoidal activation which offers us as a result the probability that if a case is positive or negative