Update src/streamlit_app.py

src/streamlit_app.py

@@ -208,11 +208,11 @@ st.markdown("""
 # ==========================================================================
 # Sidebar Configuration
 # ==========================================================================
-st.sidebar.markdown("### ⚙️ Engine Settings")
+st.sidebar.markdown("###  Engine Settings")
 top_k_slider = st.sidebar.slider("Number of results (K)", min_value=3, max_value=20, value=6, step=1)
 
 st.sidebar.markdown("---")
-st.sidebar.markdown("### 🏆 Trained Backbone Model")
+st.sidebar.markdown("### Trained Backbone Model")
 st.sidebar.markdown("""
 * **Architecture**: ResNet-50 (CIFAR STEM adjusted)
 * **Training Type**: Self-Supervised (SimCLR)
@@ -224,14 +224,14 @@ st.sidebar.markdown("""
 """)
 
 st.sidebar.markdown("---")
-st.sidebar.markdown("### ⚡ Backend Pipeline")
+st.sidebar.markdown("###  Backend Pipeline")
 st.sidebar.markdown("""
 * **Inference Engine**: ONNX Runtime CPU
 * **Vector Database**: FAISS (IndexFlatIP)
 * **Similarity Metric**: Exact Cosine Similarity
 * **Database Size**: 10,000 Images
 """)
-st.sidebar.caption("Group 20, CISC 867, Queen's University, Spring 2026")
+st.sidebar.caption("Mahmoud Alyosify - Natalie Monged & Mirna Embaby, CISC 867, Queen's University, Spring 2026")
 
 # ==========================================================================
 # Main Title
@@ -240,7 +240,7 @@ st.title("Real-time SimCLR Image Retrieval Engine")
 st.markdown("##### Self-Supervised Representation Learning with ResNet-50 & FAISS Indexing")
 
 # Define Tabs
-tab1, tab2 = st.tabs(["🔍 Real-Time Search", "📊 Ablation Study Dashboard"])
+tab1, tab2 = st.tabs(["Real-Time Search", "Ablation Study Dashboard"])
 
 # ==========================================================================
 # TAB 1: Visual Search Engine
@@ -256,7 +256,7 @@ with tab1:
     query_info = None
 
     with col_left:
-        st.markdown("### 📤 Select Query Image")
+        st.markdown("### Select Query Image")
         
         # Method 1: Upload a file
         uploaded_file = st.file_uploader("Upload an image...", type=["jpg", "png", "jpeg"])
@@ -289,7 +289,7 @@ with tab1:
                 st.error("Reference image file not found.")
 
     with col_right:
-        st.markdown("### 🔎 Visual Similarity Search Results")
+        st.markdown("### Visual Similarity Search Results")
         
         if query_image is not None:
             t_start = time.time()
@@ -371,7 +371,7 @@ with tab1:
                     col.markdown(card_content, unsafe_allow_html=True)
                     
         else:
-            st.info("💡 Please upload a custom image or click the button to select a random test image to query the visual search engine!")
+            st.info("Please upload a custom image or click the button to select a random test image to query the visual search engine!")
 
 # ==========================================================================
 # TAB 2: Ablation Study Dashboard
@@ -412,7 +412,7 @@ with tab2:
             st.markdown(card_html, unsafe_allow_html=True)
             
     st.markdown("---")
-    st.markdown("#### 💡 Key Project Takeaways")
+    st.markdown("####  Key Project Takeaways")
     col1, col2 = st.columns(2)
     
     with col1:
@@ -420,16 +420,16 @@ with tab2:
         ##### 1. The Shortcut-Learning Problem 🧩
         Without Color Jitter, contrastive self-supervised encoders exploit low-level **color histograms** as a shortcut to maximize mutual information, rather than learning general shapes and semantic features. This results in weaker downstream representations, showing a severe performance limit (e.g. baseline Crop+Flip+Blur achieves only **64.49%**).
         
-        ##### 2. The Color Jitter Shield 🛡️
+        ##### 2. The Color Jitter Shield 
         Adding photometric distortion (color jittering) forces the model to ignore color profiles and focus on invariant structures, spatial boundaries, and contours. This single ablation yields a massive average boost of **+18.1 pp** across all settings, pushing our best encoder (Exp 41) to a stellar **84.30% Top-1 accuracy**!
         """)
         
     with col2:
         st.markdown("""
-        ##### 3. Model Architecture & Stem Tuning ⚙️
+        ##### 3. Model Architecture & Stem Tuning 
         Modifying the standard ResNet-50 conv1 stem from 3x3 (stride 1) and removing the initial MaxPool was crucial to preserve the resolution of 32x32 CIFAR-10 images. 
         
-        ##### 4. Near Foundation-Model Upper Bound 🏆
+        ##### 4. Near Foundation-Model Upper Bound 
         Our zero-shot CLIP ViT-B/32 foundation model evaluation sets the academic upper bound at **88.80%**. Our custom-trained SimCLR ResNet-50 achieves **95% of this performance** (**84.30%**) while using **8,000x less training data**!
         """)