# Function to preprocess image for the model
# def preprocess_image(img_path):
# img = Image.open(img_path) # target_size=(HEIGHT, WIDTH)
# img_array = np.asarray(img.resize((IMG_SIZE,) * 2, resample=Image.LANCZOS))
# # img_array /= 255.0 # Normalize
# img_array = img_array[..., :3] / 255.0
# img_array = np.expand_dims(img_array, axis=0)
# img_array = load_ben_color(img)
# # image = np.asarray(image.resize(shape))
# # image = np.array(image)[..., :3] / 255.0
# # image = np.expand_dims(image, axis=0)
# return np.asarray(img_array)
# Function to load BEN color
# def load_ben_color(file):
# img = cv2.imread(file)
# # gray = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
# img = cv2.resize(file, (IMG_SIZE, IMG_SIZE))
# img = cv2.addWeighted(img, 4, cv2.GaussianBlur(img, (0, 0), sigmaX), -4, 128)
# return img
# img = preprocess_image(file_path)
# prediction = classifier_model.predict(img, 2)
// type: "image/jpeg"
// var reader = new FileReader();
// reader.onload = function (e) {
// var base64data = e.target.result; // Base64 encoded image data
// $.ajax({
// url: "/api/v1/image-classify",
// type: "POST",
// data: {
// file: base64data, // Submit encoded data
// },
// success: function (data) {
// // Handle successful upload response
// alert("Image uploaded successfully!");
// },
// error: function (error) {
// // Handle upload error
// alert("Error uploading image: " + error.responseText);
// }
// });
// };
// reader.readAsDataURL(file); // Read image as Base64
// const form = new FormData();
// form.append("file", "/home/fortesenselabs/Tech/seekersoftec/Dr. Odigie-wares/datasets/DR dataset/1.jpg");
// const settings = {
// "async": true,
// "crossDomain": true,
// "url": "http://localhost:8000/api/v1/image-classify",
// "method": "POST",
// "headers": {},
// "processData": true,
// "contentType": false,
// "mimeType": "multipart/form-data",
// "data": form
// };
// $.ajax(settings).done(function (response) {
// console.log(response);
// });